window.LOCAL_FRONTIER_PROFILE_DATA = { "schema_version": "1.0.0", "generated_at": "2026-07-09T00:00:00.000Z", "profiles": [ { "id": "abhishekchohan--gemma-3-12b-it-quantized-w4a16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "abhishekchohan/gemma-3-12b-it-quantized-W4A16", "title": "abhishekchohan Gemma 3 12B IT W4A16", "summary": "Audited memory-side text-decode bounds profile for the abhishekchohan compressed-tensors W4A16 Gemma 3 12B IT serving artifact.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-12b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata and the served compressed-tensors config", "config_compatible": false, "notes": "The model card/API metadata identify google/gemma-3-12b-it as the base model, but that base repo remains gated in this audit environment. This profile therefore uses the public quantized derivative config and safetensors headers directly instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-12b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 12.271375632, "swept_params_b": 11.850084768, "auxiliary_resident_params_b": 0.421290864, "resident_weight_gb": 8.405170656, "swept_weight_gb": 7.562588928, "auxiliary_resident_weight_gb": 0.842581728, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "ordinary text decode charges language_model tensors including the tied input embedding/output projection because no separate lm_head.weight is stored", "auxiliary_scope": "vision_tower and multi_modal_projector tensors are resident for the multimodal package but not swept as full matrices for each generated text token", "notes": "Range-read safetensors headers record 1737 tensors totaling 12.271375632B logical parameters and 8.405170656 GB payload bytes. The checkpoint stores packed I32 int4 tensors, BF16 unquantized tensors, and tiny I64 shape tensors. The header has no language_model.lm_head.weight tensor, so language_model.model.embed_tokens.weight is charged as the tied output projection for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 8, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window_pattern 6 over 48 language layers. Using the documented Gemma 3 pattern of five local layers followed by one global layer gives 8 full-context global layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 language layers use the config's 1024-token local sliding-window attention." } ], "notes": "Layered KV models ordinary text decode after any image prefill. The config records cache_implementation hybrid and kv_cache_scheme null, so the profile charges BF16 K and V streams." }, "notes": "Gemma3ForConditionalGeneration is multimodal. This profile models ordinary text decode, not vision encoder or image-prefill throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6849411922557306, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-gemma3-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 ignored modules and scale tensors, and I64 shape side tensors from safetensors headers. W4A16 dequantization, activation traffic, compute throughput, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The model card describes 4-bit weight quantization and 16-bit activations. The served config records compressed-tensors pack-quantized int4 weights with group size 128, bfloat16 model dtype, and kv_cache_scheme null; this profile therefore charges exact stored tensor bytes for weights and BF16 KV cache traffic." }, "evidence": [ { "label": "abhishekchohan Gemma 3 12B W4A16 model card and API metadata", "url": "https://huggingface.co/api/models/abhishekchohan/gemma-3-12b-it-quantized-W4A16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit d9833e9131576955f818012055a0bc9b74b46e64, the API reports a public non-gated Gemma-licensed image-text-to-text repo with base_model google/gemma-3-12b-it, compressed-tensors tags, region:us, 195261 downloads, and safetensors logical parameters I64: 672, I32: 10758389760, BF16: 1512985200, total: 12271375632. The model card identifies the artifact as a W4A16 quantized derivative of Google's Gemma 3 instruction-tuned models using LLM Compressor." }, { "label": "abhishekchohan Gemma 3 12B W4A16 config", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/raw/d9833e9131576955f818012055a0bc9b74b46e64/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "sliding_window_pattern", "max_context_tokens", "serving", "vision_geometry" ], "notes": "The config records Gemma3ForConditionalGeneration, gemma3_text, bfloat16 model dtype, compressed-tensors pack-quantized int4 weights, group_size 128, symmetric minmax W4A16 quantization, kv_cache_scheme null, cache_implementation hybrid, 48 text layers, 3840 hidden size, 15360 intermediate size, 16 attention heads, 8 KV heads, 256 head dimension, 131072 max positions, 1024-token sliding window, sliding_window_pattern 6, and a 27-layer SigLIP vision tower." }, { "label": "abhishekchohan Gemma 3 12B W4A16 quantization recipe", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/raw/d9833e9131576955f818012055a0bc9b74b46e64/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope", "swept_weight_gb" ], "notes": "The LLM Compressor recipe targets Linear modules with scheme W4A16 and ignores lm_head, multi_modal_projector, and vision_tower modules. Safetensors headers show no separate lm_head.weight tensor, so the tied embedding/output projection is charged in ordinary text-decode traffic." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes the repeating local/global attention pattern as five local attention layers followed by one global attention layer, and says larger Gemma 3 models support 128k context. Applied to the 48 layers in the served config, that yields 8 global layers and 40 local layers." }, { "label": "abhishekchohan Gemma 3 12B W4A16 safetensors index and headers", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/resolve/d9833e9131576955f818012055a0bc9b74b46e64/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout", "vision_scope" ], "notes": "The safetensors index records total_size 8405170656 bytes across two shards. Range-reads of both shard headers found 1737 tensors totaling the same 8.405170656 GB: I32 packed tensors total 5.379194880 GB, BF16 tensors total 3.025970400 GB, and I64 side tensors total 0.000005376 GB. Logical parameters total 12.271375632B after expanding weight_packed I32 tensors by 8. Ordinary swept text traffic totals 11.850084768B logical parameters / 7.562588928 GB: language MLP 4.379445504 GB, language self-attention 1.167903744 GB, other language layer tensors 0.001474560 GB, language norm 0.000007680 GB, and tied language_model.model.embed_tokens.weight output projection 2.013757440 GB. Resident-only tensors total 0.421290864B logical parameters / 0.842581728 GB: vision_tower 0.833732064 GB and multi_modal_projector 0.008849664 GB." }, { "label": "Google Gemma 3 12B IT gated base access note", "url": "https://huggingface.co/google/gemma-3-12b-it", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base repo remains gated in this audit environment, matching the existing google/gemma-3-12b-it unsupported profile. This audited quantized profile does not copy base geometry; it relies on the public derivative config and tensor headers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card metadata, served config, quantization recipe, safetensors index, direct safetensors shard header range reads, and Gemma 3 local/global attention documentation." }, "notes": "This profile supersedes the generated row's incorrect I32 full-precision weight estimate with exact compressed-tensors W4A16 stored bytes and a Gemma 3 layered local/global KV adapter." }, { "id": "aeon-7--qwen3-6-35b-a3b-heretic-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "title": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for the AEON-7 llm-compressor NVFP4 package of Qwen3.6 35B A3B Heretic.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "tvall43/Qwen3.6-35B-A3B-heretic", "relation": "quantized", "source": "Hugging Face base_model metadata, served config comparison, and safetensors header review", "config_compatible": true, "notes": "The AEON-7 artifact records tvall43/Qwen3.6-35B-A3B-heretic as its quantized base. That base records Qwen/Qwen3.6-35B-A3B as its finetune base. Manual comparison found the same checked text geometry across AEON-7, tvall43, and official Qwen configs: 40 text layers, full_attention_interval 4, 256 routed experts, 8 routed experts per token, shared_expert_intermediate_size 512, and 262144 max positions." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 23.336710112, "main_resident_weight_gb": 21.426448896, "auxiliary_resident_weight_gb": 1.910261216, "fixed_weight_gb": 3.306809856, "routed_expert_weight_gb": 0.07077984, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "single_safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding and vision tower tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The AEON-7 recipe leaves visual modules, embed_tokens, linear_attn modules, mlp.gate, shared_expert_gate, router/norm tensors, and lm_head outside the NVFP4 Linear target set. Routed expert tensor bytes are uniform across all 256 expert indexes. Unlike some sibling Qwen3.6 packages, this repo has no separate resident MTP safetensors sidecar; DFlash speculative decoding uses a separate drafter repo and is outside this ordinary-decode profile." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The AEON-7 NVFP4 artifact preserves the base Qwen3.6 text architecture, so quantizing weights and activations does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and DFlash speculative decoding are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower. This profile models ordinary cached text decode through the language model and output head, with resident-only multimodal tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-llm-compressor-nvfp4-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored NVFP4 packed weights, FP8 scale tensors, BF16 unquantized tensors, F32 scalar scales, and BF16 KV bytes. Activation quantization, dequantization, compute overhead, write traffic, and DFlash acceptance are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights and activations with kv_cache_scheme null. The AEON-7 model card says not to set --kv-cache-dtype for the DFlash serving path because the non-causal DFlash drafter requires BF16 KV, so this profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4 API metadata", "url": "https://huggingface.co/api/models/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b", "serving" ], "notes": "At commit 3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd, the API reports a public non-gated Apache-2.0 Transformers image-text-to-text repo with base_model tvall43/Qwen3.6-35B-A3B-heretic, safetensors dtype counts F32 61760, BF16 2496468336, F8_E4M3 2038169600, U8 16305356800, and total 20840056496 logical tensor elements. Tags include qwen3_5_moe, compressed-tensors, nvfp4, fp4, vllm, dgx-spark, dflash, fp8-kv-cache, and region:us. Current API downloads were 203785 when audited." }, { "label": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4 model card", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "source_type": "model_card", "supports": [ "serving", "runtime_format", "kv_store_format", "speculative_decode_scope" ], "notes": "The card describes the repo as a v2 multimodal-preserved NVFP4 quantization using llmcompressor compressed-tensors and vLLM. The production commands use vLLM with --quantization compressed-tensors, --attention-backend flash_attn, --enable-chunked-prefill, --enable-prefix-caching, and a separate DFlash drafter repo. The recipe notes explicitly say not to set --kv-cache-dtype for the DFlash path because BF16 KV is required." }, { "label": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4 config", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4/raw/3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "quantization_ignore_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, tie_word_embeddings false, a vision_config, and kv_cache_scheme null. The compressed-tensors config targets Linear with nvfp4-pack-quantized format and a 341-entry ignore list covering visual modules, linear_attn projections, mlp.gate, shared_expert_gate, and lm_head." }, { "label": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4 quantization recipe", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4/raw/3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantization_ignore_scope" ], "notes": "The recipe targets Linear with scheme NVFP4 and ignores lm_head, embed_tokens, visual modules, vision_model, mm_projector, mlp.gate, shared_expert_gate, mlp.router, linear_attn modules, norms, q_norm, and k_norm." }, { "label": "tvall43 Qwen3.6 35B A3B Heretic base config", "url": "https://huggingface.co/tvall43/Qwen3.6-35B-A3B-heretic/raw/9b10fb891b6dd485ead5a4da1c324e7527e4df69/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison found the checked Qwen3.6 text and vision geometry preserved between the unquantized Heretic base and the AEON-7 NVFP4 artifact. The Heretic base metadata records Qwen/Qwen3.6-35B-A3B as its finetune base." }, { "label": "Qwen3.6 35B A3B official base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison against the official base config found the same checked text geometry used by the AEON-7 artifact and the tvall43 Heretic base." }, { "label": "AEON-7 Qwen3.6 35B A3B Heretic NVFP4 safetensors header", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4/resolve/3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Direct range-read of the single safetensors header found 124306 tensors with tensor data totaling 23.336710112 GB. Stored bytes split into F32 0.00024704 GB, BF16 4.992936672 GB, F8_E4M3 2.0381696 GB, and U8 16.3053568 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.426448896 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1.910261216 GB. Main routed expert tensors sum to 18.11963904 GB, or 0.07077984 GB per expert index. Fixed ordinary text traffic sums to 3.306809856 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current Hugging Face API metadata, pinned model card, served NVFP4 config, quantization recipe, tvall43 Heretic base config, official Qwen base config, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile intentionally models ordinary text decode for the AEON-7 NVFP4 body only. The model card's higher DGX Spark throughput numbers use a separate DFlash drafter with speculative decoding; those gains require a future speculative-decoding profile and should not be folded into this memory-side body profile." }, { "id": "allenai--molmo2-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/Molmo2-8B", "title": "AllenAI Molmo2 8B F32", "summary": "Audited memory-side text-decode bounds profile for the F32 Molmo2 8B multimodal repo.", "model_family": "molmo2-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B + google/siglip-so400m-patch14-384", "relation": "finetune", "source": "Hugging Face model metadata, served Molmo2 config, and safetensors header layout", "config_compatible": false, "notes": "The model metadata identifies Qwen/Qwen3-8B and google/siglip-so400m-patch14-384 as bases. The served checkpoint is a custom Molmo2 image-text-to-text wrapper with Qwen3-like text geometry, custom tensor names, F32 storage, a 36864-token text context, and a SigLIP-derived vision backbone, so this profile uses the served Molmo2 config directly." }, "architecture": { "canonical_architecture_id": "molmo2-8b", "max_context_tokens": 36864, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.66170312, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 1.093297616, "resident_weight_gb": 34.64681248, "swept_weight_gb": 30.273622016, "auxiliary_resident_weight_gb": 4.373190464, "resident_parameter_scope": "safetensors_header_stored_f32_full_text_vision_package", "swept_parameter_scope": "model.transformer.blocks plus model.transformer.ln_f plus lm_head.weight", "auxiliary_scope": "model.transformer.wte input embeddings plus model.vision_backbone image_vit, image_projector, and image_pooling_2d tensors", "notes": "Range-read safetensors headers record 706 F32 tensors totaling 8661703120 stored parameters and 34.64681248 GB. Ordinary text decode sweeps the text decoder blocks, final norm, and the separate lm_head output projection. The input embeddings, additional 128-token embedding, and vision backbone are resident for the multimodal package but are not swept as full matrices for each generated text token." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text_config records 36 text layers, 8 KV heads, 128 head dimension, no sliding-window setting, and use_cache true. Because the served config dtype is float32 and all stored tensors are F32, this profile charges full-context F32 K and V streams for text decode." }, "notes": "Molmo2ForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill and does not estimate vision encoder or image preprocessing throughput." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "fp32", "kv_store_bytes_per_scalar": 4, "kv_read_format": "fp32", "kv_read_bytes_per_scalar": 4, "runtime_format": "transformers-fp32-molmo2-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges F32 stored tensors and F32 KV streams from the served config and tensor headers.", "notes": "The repo config records dtype float32, and range-read safetensors headers record only F32 tensors." }, "evidence": [ { "label": "Molmo2 8B model card", "url": "https://huggingface.co/allenai/Molmo2-8B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model_proof" ], "notes": "The current Hugging Face API metadata records a public non-gated Apache-2.0 image-text-to-text repo with transformers, custom_code, region:us, base_model Qwen/Qwen3-8B, and base model card data for Qwen/Qwen3-8B plus google/siglip-so400m-patch14-384." }, { "label": "Molmo2 8B config", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "vision_backbone_scope", "tie_word_embeddings" ], "notes": "At commit e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b, the config records Molmo2ForConditionalGeneration, dtype float32, tie_word_embeddings false, a text_config with 36 layers, 4096 hidden size, 32 attention heads, 8 KV heads, 128 head dimension, 36864 max positions, 151936 vocab, and 128 additional vocab entries, plus adapter and ViT configs for the multimodal package." }, { "label": "Molmo2 8B custom modeling code", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/modeling_molmo2.py", "source_type": "manual_review", "supports": [ "ordinary_text_decode_scope", "kv_adapter", "lm_head_layout" ], "notes": "Manual review found Molmo2CausalLMOutputWithPast, DynamicCache use, past_key_values updates in attention, Molmo2ForConditionalGeneration with a separate lm_head, and generation support. This supports modeling cached causal text decode separately from vision prefill." }, { "label": "Molmo2 8B Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/Molmo2-8B", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "The current API response records commit e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b, 603471 downloads when audited, safetensors parameters F32: 8661703120 and total: 8661703120, and usedStorage 34660269133 bytes." }, { "label": "Molmo2 8B safetensors headers", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_parameters 8661703120 and total_size 34646812480 across eight shards. Range-read shard headers found 706 F32 tensors. model.transformer.blocks total 27.784286208 GB, model.transformer.ln_f totals 0.000016384 GB, lm_head.weight is a separate [151936, 4096] F32 tensor of 2.489319424 GB, model.transformer.wte embeddings total 2.491416576 GB, and vision_backbone tensors total 1.881773888 GB." }, { "label": "Qwen3 8B profile comparison", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The text geometry is Qwen3-like, but the served Molmo2 config changes context, tensor names, dtype, and multimodal wrapping, so the profile does not copy the existing Qwen3 8B profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served config, pinned custom modeling code, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the full resident package as swept traffic and undercounted F32 KV by using BF16-sized KV coefficients. It is for ordinary text decode bounds after any multimodal prefill." }, { "id": "allenai--molmo2-o-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/Molmo2-O-7B", "title": "AllenAI Molmo2-O 7B F32", "summary": "Audited memory-side text-decode bounds profile for the F32 Molmo2-O 7B multimodal repo.", "model_family": "molmo2-dense", "base_model_proof": { "base_model": "allenai/Olmo-3-7B-Instruct + google/siglip-so400m-patch14-384", "relation": "finetune", "source": "Hugging Face model metadata, served Molmo2 config, custom modeling code, and safetensors header layout", "config_compatible": false, "notes": "The repo metadata identifies allenai/Olmo-3-7B-Instruct and google/siglip-so400m-patch14-384 as bases. The served checkpoint is a custom Molmo2 image-text-to-text wrapper with Olmo-3-like text geometry, custom tensor names, F32 storage, a 65536-token text context, and a SigLIP-derived vision backbone, so this profile uses the served Molmo2 config directly." }, "architecture": { "canonical_architecture_id": "molmo2-o-7b", "max_context_tokens": 65536, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.760786896, "swept_params_b": 6.887272448, "auxiliary_resident_params_b": 0.873514448, "resident_weight_gb": 31.043147584, "swept_weight_gb": 27.549089792, "auxiliary_resident_weight_gb": 3.494057792, "resident_parameter_scope": "safetensors_header_stored_f32_full_text_vision_package", "swept_parameter_scope": "model.transformer.blocks plus model.transformer.ln_f plus lm_head.weight", "auxiliary_scope": "model.transformer.wte input embeddings plus model.vision_backbone tensors", "notes": "Range-read safetensors headers record 674 F32 tensors totaling 7760786896 stored parameters and 31.043147584 GB. Ordinary text decode sweeps the text decoder blocks, final norm, and the separate lm_head output projection. The input embeddings, additional 128-token embedding, and vision backbone are resident for the multimodal package but are not swept as full matrices for each generated text token." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text_config records 32 text layers, 32 KV heads, 128 head dimension, 65536 max positions, and use_cache true. Manual review of the custom modeling code found standard DynamicCache updates in every Molmo2Attention layer and no sliding-window or local-attention cache truncation. Because the served config dtype is float32 and all stored tensors are F32, this profile charges full-context F32 K and V streams for text decode." }, "notes": "Molmo2ForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill and does not estimate vision encoder, image projector, image pooling, or preprocessing throughput." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "fp32", "kv_store_bytes_per_scalar": 4, "kv_read_format": "fp32", "kv_read_bytes_per_scalar": 4, "runtime_format": "transformers-fp32-molmo2-olmo3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges F32 stored tensors and F32 KV streams from the served config and tensor headers.", "notes": "The repo config records dtype float32, and range-read safetensors headers record only F32 tensors." }, "evidence": [ { "label": "Molmo2-O 7B model card", "url": "https://huggingface.co/allenai/Molmo2-O-7B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model_proof" ], "notes": "The current Hugging Face API metadata records a public non-gated Apache-2.0 image-text-to-text repo with transformers, safetensors, molmo2, olmo, multimodal, custom_code, region:us, base_model allenai/Olmo-3-7B-Instruct, and card data for allenai/Olmo-3-7B-Instruct plus google/siglip-so400m-patch14-384." }, { "label": "Molmo2-O 7B Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/Molmo2-O-7B", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "The current API response records commit 784410650d12be9bc086118fdefa32d2c3bced86, 161905 downloads when audited, safetensors parameters F32: 7760786896 and total: 7760786896, and usedStorage 31043241424 bytes." }, { "label": "Molmo2-O 7B config", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "vision_backbone_scope", "tie_word_embeddings" ], "notes": "At commit 784410650d12be9bc086118fdefa32d2c3bced86, the config records Molmo2ForConditionalGeneration, dtype float32, tie_word_embeddings false, a text_config with 32 layers, 4096 hidden size, 32 attention heads, 32 KV heads, 128 head dimension, 65536 max positions, 100278 vocab, 128 additional vocab entries, YaRN RoPE scaling for selected layers, and adapter/ViT settings for the multimodal package." }, { "label": "Molmo2-O 7B custom modeling code", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/modeling_molmo2.py", "source_type": "manual_review", "supports": [ "ordinary_text_decode_scope", "kv_adapter", "lm_head_layout" ], "notes": "Manual review found Molmo2CausalLMOutputWithPast, DynamicCache use, past_key_values updates in attention, Molmo2ForConditionalGeneration with a separate lm_head, and generation support. The code varies RoPE scaling by layer but does not implement sliding-window or local-attention KV truncation, so ordinary cached text decode is modeled separately from vision prefill with full-context KV." }, { "label": "Molmo2-O 7B safetensors index and shard headers", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_parameters 7760786896 and total_size 31043147584 bytes across seven shards. Range-read shard headers found 674 F32 tensors. model.transformer.blocks total 25.906118656 GB, model.transformer.ln_f totals 0.000016384 GB, lm_head.weight is a separate [100278, 4096] F32 tensor of 1.642954752 GB, model.transformer.wte embeddings total 1.645051904 GB, and vision_backbone tensors total 1.849005888 GB. Linked-object HEAD checks resolved the seven shards to 31.043241424 GB total, leaving 93840 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "OLMo 3 7B Instruct profile comparison", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "The text geometry is Olmo-3-like, but the served Molmo2 config changes tensor names, dtype, multimodal wrapping, and cache behavior evidence. The Molmo2 custom code lacks the OLMo3 layer_types sliding-window pattern, so this profile does not copy the existing OLMo3 hybrid-KV profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned served config, pinned custom modeling code, direct range-read safetensors shard headers, linked-object HEAD checks, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate, which charged the full resident package as active traffic while undercounting F32 KV by using half-sized KV coefficients. It is for ordinary text decode bounds after any multimodal prefill." }, { "id": "allenai--olmo-2-0425-1b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/OLMo-2-0425-1B", "title": "AllenAI OLMo 2 0425 1B F32", "summary": "Audited memory-side text-decode bounds profile for the F32 OLMo 2 0425 1B base checkpoint.", "model_family": "olmo2-dense", "base_model_proof": { "base_model": "allenai/OLMo-2-0425-1B", "relation": "base", "source": "Hugging Face model card, served config, API metadata, and direct safetensors header metadata", "config_compatible": true, "notes": "This repo is the OLMo 2 1B base checkpoint. The profile uses the served config directly rather than deriving architecture from the model name." }, "architecture": { "canonical_architecture_id": "olmo-2-0425-1b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.484916736, "swept_params_b": 1.27939584, "auxiliary_resident_params_b": 0.205520896, "resident_weight_gb": 5.939666944, "swept_weight_gb": 5.11758336, "auxiliary_resident_weight_gb": 0.822083584, "resident_parameter_scope": "safetensors_header_stored_f32", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes the separate lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 179 F32 tensors totaling 1.484916736B stored parameters / 5.939666944 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges resident bytes for both tables, excludes only the input embedding lookup from ordinary decode traffic, and keeps lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 16 layers, 16 attention heads, 16 KV heads, hidden size 2048, no sliding-window setting, and use_cache true, so Bounds Engine v1 charges full-context K and V streams for every layer." }, "notes": "Dense Olmo2ForCausalLM profile using the served config and direct safetensors header grouping." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "fp32", "kv_store_bytes_per_scalar": 4, "kv_read_format": "fp32", "kv_read_bytes_per_scalar": 4, "runtime_format": "transformers-fp32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges F32 stored tensors and F32 KV streams from the served config and tensor headers.", "notes": "The repo config records torch_dtype float32, and direct safetensors headers record only F32 tensors." }, "evidence": [ { "label": "OLMo 2 0425 1B model card", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model card identifies this as OLMo 2 1B, the smallest OLMo 2 base model, with 16 layers, hidden size 2048, 16 attention heads, 4096 context length, Apache-2.0 licensing, and standard Transformers inference." }, { "label": "OLMo 2 0425 1B HF API metadata", "url": "https://huggingface.co/api/models/allenai/OLMo-2-0425-1B", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "At commit a1847dff35000b4271fa70afc5db10fd29fedbdf, the API records a public non-gated Apache-2.0 text-generation Transformers repo with olmo2, safetensors, endpoints_compatible, region:us, 202403 downloads, and safetensors parameters F32: 1484916736, total: 1484916736." }, { "label": "OLMo 2 0425 1B config", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B/raw/a1847dff35000b4271fa70afc5db10fd29fedbdf/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Olmo2ForCausalLM, torch_dtype float32, tie_word_embeddings false, hidden size 2048, intermediate size 8192, 16 layers, 16 attention heads, 16 KV heads, 4096 max position embeddings, rope_theta 500000, vocab size 100352, and use_cache true." }, { "label": "OLMo 2 0425 1B safetensors index and shard headers", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B/raw/a1847dff35000b4271fa70afc5db10fd29fedbdf/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_size 5939666944 bytes across two shards. Range-read shard headers found 179 F32 tensors totaling 5.939666944 GB: shard 1 has 173 tensors / 4.983341056 GB and shard 2 has 6 tensors / 0.956325888 GB. model.embed_tokens.weight has shape [100352, 2048] and contributes 0.822083584 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Decoder layers plus model.norm.weight plus lm_head.weight total 5.117583360 GB. Linked-object HEAD checks resolved the two shards to 5.939687552 GB total, leaving 20608 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served config, model card, safetensors index, linked-object HEAD checks, direct range-read safetensors shard headers, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate. The generated row correctly detected F32 weights but used BF16-sized KV coefficients; this audited profile charges FP32 KV because the served config and all safetensors payloads are float32." }, { "id": "allenai--olmo-3-7b-instruct-sft", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/Olmo-3-7B-Instruct-SFT", "title": "AllenAI OLMo 3 7B Instruct SFT BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 OLMo 3 7B Instruct SFT checkpoint.", "model_family": "olmo3-dense", "base_model_proof": { "base_model": "allenai/Olmo-3-1025-7B", "relation": "finetune", "source": "Hugging Face model card base_model metadata, served SFT config, base-model config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The repo metadata identifies allenai/Olmo-3-1025-7B as the base model. Manual comparison found no differences in checked architecture fields between the SFT config and the base config, so this profile uses the served SFT config directly." }, "architecture": { "canonical_architecture_id": "olmo3-7b", "max_context_tokens": 65536, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.298011136, "swept_params_b": 6.887272448, "auxiliary_resident_params_b": 0.410738688, "resident_weight_gb": 14.596022272, "swept_weight_gb": 13.774544896, "auxiliary_resident_weight_gb": 0.821477376, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes the separate lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 355 BF16 tensors totaling 7.298011136B stored parameters / 14.596022272 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges resident bytes for both tables, excludes only the input embedding lookup from ordinary decode traffic, and keeps lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config layer_types marks layers 3, 7, 11, 15, 19, 23, 27, and 31 as full_attention. These layers retain full-context K and V cache." }, { "kind": "sliding_window", "layers": 24, "kv_heads": 32, "head_dim": 128, "window_tokens": 4096, "kv_scalar_multiplier": 2, "notes": "The remaining 24 layers are sliding_attention with a 4096-token window." } ], "notes": "Transformers 4.57.1 Olmo3Attention sets sliding_window only for sliding_attention layers and the model builds separate full_attention and sliding_attention causal masks. Bounds Engine v1 represents this as explicit full-context plus sliding-window KV components." }, "notes": "Dense Olmo3ForCausalLM profile using the served SFT config, Transformers OLMo3 attention implementation, and direct safetensors header grouping." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-olmo3-hybrid-attention-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges BF16 stored tensors and BF16 KV streams from the served config and tensor headers.", "notes": "The repo config records dtype bfloat16, and direct safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "OLMo 3 7B Instruct SFT model card", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model card identifies this as the Olmo 3 7B Instruct SFT stage, records Apache-2.0 licensing, describes Olmo 3 as a 7B/32B model family, and links the SFT row to base model allenai/Olmo-3-1025-7B." }, { "label": "OLMo 3 7B Instruct SFT HF API metadata", "url": "https://huggingface.co/api/models/allenai/Olmo-3-7B-Instruct-SFT", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit e1452fc572d51966ff4aaeb25118b891eb93e549, the API records a public non-gated Apache-2.0 text-generation Transformers repo with olmo3, safetensors, endpoints_compatible, region:us, base_model allenai/Olmo-3-1025-7B, 214141 downloads, and safetensors parameters BF16: 7298011136, total: 7298011136." }, { "label": "OLMo 3 7B Instruct SFT config", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT/raw/e1452fc572d51966ff4aaeb25118b891eb93e549/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Olmo3ForCausalLM, dtype bfloat16, tie_word_embeddings false, hidden size 4096, intermediate size 11008, 32 layers, 32 attention heads, 32 KV heads, 65536 max position embeddings, 4096 sliding window, a layer_types pattern with 24 sliding_attention layers and every fourth layer as full_attention, YaRN RoPE scaling factor 8 from original 8192 positions, rope_theta 500000, vocab size 100278, attention_bias false, and use_cache true." }, { "label": "OLMo 3 7B base config comparison", "url": "https://huggingface.co/allenai/Olmo-3-1025-7B/raw/a81bae42db3975be1671e27b9c9a56da1a9f980f/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences between the SFT config and the base config in checked architecture fields: model class, model type, dtype, tied embeddings, hidden size, intermediate size, layer count, attention heads, KV heads, max context, sliding window, layer_types, RoPE settings, vocabulary size, attention_bias, and use_cache." }, { "label": "Transformers 4.57.1 OLMo3 implementation", "url": "https://github.com/huggingface/transformers/blob/v4.57.1/src/transformers/models/olmo3/modeling_olmo3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "serving" ], "notes": "Manual review found Olmo3Attention reads config.layer_types for each layer, sets self.sliding_window only when attention_type is sliding_attention, updates standard K and V tensors in past_key_values, and passes either a full_attention or sliding_attention mask to each decoder layer." }, { "label": "OLMo 3 7B Instruct SFT safetensors index and shard headers", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT/raw/e1452fc572d51966ff4aaeb25118b891eb93e549/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_parameters 7298011136 and total_size 14596022272 bytes across three shards. Range-read shard headers found 355 BF16 tensors totaling 14.596022272 GB, matching the index total_size: shard payloads are 4.969971712 GB, 4.981145600 GB, and 4.644904960 GB. model.embed_tokens.weight has shape [100278, 4096] and contributes 0.821477376 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Decoder layers plus model.norm.weight plus lm_head.weight total 13.774544896 GB. Linked-object HEAD checks resolved the three shards to 14.596063712 GB total, leaving 41440 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served config, base config comparison, model card, Transformers OLMo3 implementation, linked-object HEAD checks, safetensors index, direct range-read safetensors shard headers, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate. The generated row treated all 32 layers as full-context KV at every context length; this audited profile charges 8 full-context layers plus 24 sliding-window layers capped at 4096 tokens." }, { "id": "allenai--olmo-3-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/Olmo-3-7B-Instruct", "title": "AllenAI OLMo 3 7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 OLMo 3 7B Instruct final checkpoint.", "model_family": "olmo3-dense", "base_model_proof": { "base_model": "allenai/Olmo-3-7B-Instruct-DPO", "relation": "finetune", "source": "Hugging Face model card base_model metadata, served final config, DPO and SFT config comparisons, model card stage table, and direct safetensors header metadata", "config_compatible": true, "notes": "The repo metadata identifies allenai/Olmo-3-7B-Instruct-DPO as the immediate base model, and the model card stage table identifies this final model as the RLVR stage after SFT and DPO. Manual comparison found no differences in checked architecture fields between the final, DPO, and SFT configs except the final config records use_cache false while DPO/SFT record true. This profile uses the served final config for architecture and treats use_cache as a runtime serving flag rather than a structural difference." }, "architecture": { "canonical_architecture_id": "olmo3-7b", "max_context_tokens": 65536, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.298011136, "swept_params_b": 6.887272448, "auxiliary_resident_params_b": 0.410738688, "resident_weight_gb": 14.596022272, "swept_weight_gb": 13.774544896, "auxiliary_resident_weight_gb": 0.821477376, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes the separate lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 355 BF16 tensors totaling 7.298011136B stored parameters / 14.596022272 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges resident bytes for both tables, excludes only the input embedding lookup from ordinary decode traffic, and keeps lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config layer_types marks layers 3, 7, 11, 15, 19, 23, 27, and 31 as full_attention. These layers retain full-context K and V cache." }, { "kind": "sliding_window", "layers": 24, "kv_heads": 32, "head_dim": 128, "window_tokens": 4096, "kv_scalar_multiplier": 2, "notes": "The remaining 24 layers are sliding_attention with a 4096-token window." } ], "notes": "Transformers 4.57.1 Olmo3Attention sets sliding_window only for sliding_attention layers and the model builds separate full_attention and sliding_attention causal masks. The final repo config records use_cache false, but the architecture and implementation still define standard K/V cache geometry; Bounds Engine v1 models ordinary KV-cache serving." }, "notes": "Dense Olmo3ForCausalLM profile using the served final config, Transformers OLMo3 attention implementation, and direct safetensors header grouping." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-olmo3-hybrid-attention-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges BF16 stored tensors and BF16 KV streams from the served config and tensor headers.", "notes": "The repo config records dtype bfloat16, and direct safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "OLMo 3 7B Instruct model card", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model card identifies this as the Olmo 3 7B Instruct final model, records Apache-2.0 licensing, links the immediate base_model metadata to allenai/Olmo-3-7B-Instruct-DPO, and shows the final model as the RLVR stage after the SFT and DPO rows." }, { "label": "OLMo 3 7B Instruct HF API metadata", "url": "https://huggingface.co/api/models/allenai/Olmo-3-7B-Instruct", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit 6e5971d9eba42665f5bd5a0fcf047f299ce1dccc, the API records a public non-gated Apache-2.0 text-generation Transformers repo with olmo3, safetensors, endpoints_compatible, deploy:azure, region:us, base_model allenai/Olmo-3-7B-Instruct-DPO, 144100 downloads, and safetensors parameters BF16: 7298011136, total: 7298011136." }, { "label": "OLMo 3 7B Instruct config", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/raw/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Olmo3ForCausalLM, dtype bfloat16, tie_word_embeddings false, hidden size 4096, intermediate size 11008, 32 layers, 32 attention heads, 32 KV heads, 65536 max position embeddings, 4096 sliding window, a layer_types pattern with 24 sliding_attention layers and every fourth layer as full_attention, YaRN RoPE scaling factor 8 from original 8192 positions, rope_theta 500000, vocab size 100278, attention_bias false, and use_cache false." }, { "label": "OLMo 3 7B DPO and SFT config comparison", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-DPO/raw/b33130b7de49f0c2553b5c2b3bc8409ff3e627d1/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found the final, DPO, and SFT configs match on checked architecture fields: model class, model type, dtype, tied embeddings, hidden size, intermediate size, layer count, attention heads, KV heads, max context, sliding window, layer_types, RoPE settings, vocabulary size, and attention_bias. The only checked difference is the use_cache flag: false in the final config and true in DPO/SFT." }, { "label": "Transformers 4.57.1 OLMo3 implementation", "url": "https://github.com/huggingface/transformers/blob/v4.57.1/src/transformers/models/olmo3/modeling_olmo3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "serving" ], "notes": "Manual review found Olmo3Attention reads config.layer_types for each layer, sets self.sliding_window only when attention_type is sliding_attention, updates standard K and V tensors in past_key_values, and passes either a full_attention or sliding_attention mask to each decoder layer." }, { "label": "OLMo 3 7B Instruct safetensors index and shard headers", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/raw/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_parameters 7298011136 and total_size 14596022272 bytes across three shards. Range-read shard headers found 355 BF16 tensors totaling 14.596022272 GB, matching the index total_size: shard payloads are 4.969971712 GB, 4.981145600 GB, and 4.644904960 GB. model.embed_tokens.weight has shape [100278, 4096] and contributes 0.821477376 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Decoder layers plus model.norm.weight plus lm_head.weight total 13.774544896 GB. Linked-object HEAD checks resolved the three shards to 14.596063712 GB total, leaving 41440 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned served config, DPO and SFT config comparison, model card, Transformers OLMo3 implementation, linked-object HEAD checks, safetensors index, direct range-read safetensors shard headers, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate. The generated row treated all 32 layers as full-context KV at every context length; this audited profile charges 8 full-context layers plus 24 sliding-window layers capped at 4096 tokens." }, { "id": "allenai--olmoe-1b-7b-0125-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/OLMoE-1B-7B-0125-Instruct", "title": "AllenAI OLMoE 1B-7B 0125 Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 OLMoE 1B-7B January 2025 instruct checkpoint.", "model_family": "olmoe-moe", "base_model_proof": { "base_model": "allenai/OLMoE-1B-7B-0125-DPO", "relation": "finetune", "source": "Hugging Face API metadata, model card, served config, and direct safetensors shard-header range reads", "config_compatible": true, "notes": "The repo card and API metadata identify this checkpoint as the RLVR/instruct variant fine-tuned from allenai/OLMoE-1B-7B-0125-DPO. The checkpoint stores its own BF16 safetensors, and the served config preserves the same OLMoE geometry used by the audited 1B-7B family: 16 layers, 16 attention heads, 16 KV heads, 2048 hidden size, 64 experts, and 8 experts per token." }, "architecture": { "canonical_architecture_id": "olmoe-1b-7b-0125", "max_context_tokens": 4096, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.838323712, "main_resident_weight_gb": 13.632278528, "auxiliary_resident_weight_gb": 0.206045184, "fixed_weight_gb": 0.74737664, "routed_expert_weight_gb": 0.201326592, "routed_experts": 64, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with exact per-expert groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "All 16 layers are sparse MoE layers with 64 per-layer expert indexes and top-8 routing. Expert tensors are stored as per-expert gate/up/down matrices under model.layers.*.mlp.experts.*, and each global expert index contributes exactly 0.201326592 GB across all layers. There is no shared expert tensor group in the served config or safetensors headers." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The architecture is ordinary full-attention OLMoE: 16 layers, 16 KV heads, and 128-dimensional key/value heads. The config default sets use_cache false, but it does not define a sliding-window, recurrent, or compressed-state adapter; Bounds Engine v1 models standard cached decode with BF16 K/V traffic." }, "notes": "This profile targets the January 2025 OLMoE instruct checkpoint directly. It uses exact stored BF16 tensor bytes and treats the input embedding as resident-only for ordinary decode, while lm_head remains swept fixed traffic because embeddings are untied." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-olmoe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Bounds Engine v1 charges stored BF16 safetensors bytes, BF16 K/V cache traffic, and exact expert groups; activation traffic, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The config records torch_dtype bfloat16, and the API safetensors block plus direct shard headers record only BF16 tensors." }, "evidence": [ { "label": "OLMoE 1B-7B 0125 Instruct API metadata", "url": "https://huggingface.co/api/models/allenai/OLMoE-1B-7B-0125-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "base_model_proof", "serving", "total_params_b" ], "notes": "At commit b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e, the current API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, olmoe, conversational, en, dataset:allenai/RLVR-GSM, arXiv 2409.02060, arXiv 2411.15124, base_model:allenai/OLMoE-1B-7B-0125-DPO, base_model:finetune:allenai/OLMoE-1B-7B-0125-DPO, endpoints_compatible, and region:us tags. Current downloads are 125412. The API safetensors block records BF16 6919161856 and total 6919161856." }, { "label": "OLMoE 1B-7B 0125 Instruct model card", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e/README.md", "source_type": "model_card", "supports": [ "repo", "license", "base_model_proof", "architecture" ], "notes": "The model card describes this repo as the final RLVR/instruct variant of the January 2025 OLMoE 1B-7B line, after SFT, DPO, and RLVR post-training. It lists allenai/OLMoE-1B-7B-0125-DPO as the finetuned-from model and links the OLMoE paper plus Tulu 3 paper." }, { "label": "OLMoE 1B-7B 0125 Instruct config", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records OlmoeForCausalLM, olmoe, bfloat16, use_cache false, 16 layers, hidden size 2048, intermediate size 1024, 16 attention heads, 16 KV heads, derived head_dim 128, 64 experts, 8 experts per token, norm_topk_prob false, max_position_embeddings 4096, tie_word_embeddings false, vocab size 50304, and no shared-expert field." }, { "label": "OLMoE 1B-7B 0125 Instruct safetensors index and shard headers", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 13838323712 bytes across three shards. Direct range-read safetensors headers found 3219 BF16 tensors with payload bytes 13.838323712 GB and linked-file size 13.838721960 GB including 0.000398248 GB of safetensors header/container bytes. model.embed_tokens.weight contributes 0.206045184 GB resident-only. Ordinary text resident bytes therefore sum to 13.632278528 GB. Routed expert tensors sum to 12.884901888 GB and divide into 64 uniform expert indexes of 0.201326592 GB. Fixed ordinary text traffic, including self-attention, routers, norms, and lm_head, sums to 0.747376640 GB. No multimodal, vision, MTP, draft, or projector tensors were found." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current Hugging Face API metadata, pinned model card, served config, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated estimate with exact BF16 stored bytes and explicit uniform expected-distinct routed expert traffic for the January 2025 OLMoE instruct checkpoint." }, { "id": "allenai--olmoe-1b-7b-0924", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "allenai/OLMoE-1B-7B-0924", "title": "AllenAI OLMoE 1B-7B 0924 BF16", "summary": "Audited memory-side bounds profile for the BF16 OLMoE 1B-7B 0924 MoE pretraining repo.", "model_family": "olmoe-moe", "architecture": { "canonical_architecture_id": "olmoe-1b-7b-0924", "max_context_tokens": 4096, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.838323712, "main_resident_weight_gb": 13.632278528, "auxiliary_resident_weight_gb": 0.206045184, "fixed_weight_gb": 0.74737664, "routed_expert_weight_gb": 0.201326592, "routed_experts": 64, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with exact per-expert groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "All 16 layers are sparse MoE layers with 64 per-layer expert indexes and top-8 routing. Expert tensors are stored as per-expert gate/up/down matrices under model.layers.*.mlp.experts.*, and each global expert index contributes exactly 0.201326592 GB across all layers. There is no shared expert tensor group in the served config or safetensors headers." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_cache true, no sliding-window field, 16 layers, 16 KV heads, and 128 head dimension. Bounds Engine v1 charges expanded BF16 K/V cache for all layers." }, "notes": "This profile targets the September 2024 OLMoE pretraining checkpoint directly. It uses exact stored BF16 tensor bytes and treats the input embedding as resident-only for ordinary decode, while lm_head remains swept fixed traffic because embeddings are untied." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-olmoe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Bounds Engine v1 charges stored BF16 safetensors bytes, BF16 K/V cache traffic, and exact expert groups; activation traffic, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The config records torch_dtype bfloat16, and the API safetensors block plus direct shard headers record only BF16 tensors." }, "evidence": [ { "label": "OLMoE 1B-7B 0924 API metadata", "url": "https://huggingface.co/api/models/allenai/OLMoE-1B-7B-0924", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 6d84c48581ece794365f2b8e9cfb043c68ade9c5, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, olmoe, moe, olmo, OLMoE-mix-0924, arXiv 2409.02060, co2_eq_emissions, endpoints_compatible, and region:us tags. Current downloads are 144794. The API safetensors block records BF16 6919161856 and total 6919161856." }, { "label": "OLMoE 1B-7B 0924 model card", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924/blob/6d84c48581ece794365f2b8e9cfb043c68ade9c5/README.md", "source_type": "model_card", "supports": [ "repo", "license", "architecture", "total_params_b", "active_params_b" ], "notes": "The model card describes OLMoE-1B-7B as a Mixture-of-Experts LLM with about 1B active and 7B total parameters, released in September 2024, and the evaluation table lists 1.3B active parameters. This profile uses exact header bytes for memory traffic while preserving the published model identity." }, { "label": "OLMoE 1B-7B 0924 config", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924/raw/6d84c48581ece794365f2b8e9cfb043c68ade9c5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records OlmoeForCausalLM, olmoe, bfloat16, use_cache true, 16 layers, hidden size 2048, 16 attention heads, 16 KV heads, derived head_dim 128, 64 experts, 8 experts per token, norm_topk_prob false, max_position_embeddings 4096, tie_word_embeddings false, vocab size 50304, and no shared-expert field." }, { "label": "OLMoE 1B-7B 0924 safetensors index and shard headers", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924/raw/6d84c48581ece794365f2b8e9cfb043c68ade9c5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 13838323712 bytes across three shards. Direct range-read safetensors headers found 3219 BF16 tensors with payload bytes 13.838323712 GB. model.embed_tokens.weight contributes 0.206045184 GB resident-only. Ordinary text resident bytes therefore sum to 13.632278528 GB. Routed expert tensors sum to 12.884901888 GB and divide into 64 uniform expert indexes of 0.201326592 GB. Fixed ordinary text traffic, including self-attention, routers, norms, and lm_head, sums to 0.747376640 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face CLI/API metadata, model card, served config, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated estimate with exact BF16 stored bytes and explicit uniform expected-distinct routed expert traffic." }, { "id": "allenai--wildguard", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "allenai/wildguard", "title": "AllenAI WildGuard BF16", "summary": "Unsupported profile stub for the auto-gated BF16 WildGuard safety moderation repo.", "model_family": "mistral-dense-safety", "base_model_proof": { "base_model": "mistralai/Mistral-7B-v0.3", "relation": "finetune", "source": "WildGuard model card lineage metadata", "config_compatible": false, "notes": "The model card says WildGuard is fine-tuned from mistralai/Mistral-7B-v0.3, but the actual WildGuard config and safetensors index are gated in this audit environment. Treat this as lineage metadata, not an audited architecture comparison." }, "architecture": { "canonical_architecture_id": "wildguard-mistral-7b-v0-3", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 7.248031744, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 7248031744 BF16 safetensors parameters for this repo. KV geometry, context length, tied embedding behavior, and swept ordinary text-decode traffic are not audited because the gated config and safetensors headers are inaccessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is auto-gated. Raw config and safetensors index requests return 401, and authenticated hf download with the configured osolmaz CLI identity returns access denied because the repository requires approval.", "notes": "Do not infer Mistral layer count, KV heads, head dimension, context length, vocabulary delta, tied embedding behavior, or swept decode traffic from the base-model name or API parameter total. Replace this with an audited adapter only after direct WildGuard config and tensor-header evidence are available." }, "notes": "This profile intentionally fails closed until the gated WildGuard config and tensor layout can be audited directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic cannot be verified.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "AllenAI WildGuard API metadata", "url": "https://huggingface.co/api/models/allenai/wildguard", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "license", "pipeline", "unsupported_reason" ], "notes": "At commit cbba4823f3e8020e5a74a5e29bf85072def6f2ff, the API reports gated: auto, text-generation pipeline, Apache-2.0 license metadata, safety/moderation/classifier tags, region:us, 264728 downloads, and safetensors parameters BF16: 7248031744." }, { "label": "AllenAI WildGuard model card", "url": "https://huggingface.co/allenai/wildguard", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "unsupported_reason" ], "notes": "Authenticated hf download of README.md succeeded and the card identifies WildGuard as a 7B safety moderation model trained on WildGuardTrain, Apache-2.0 licensed, and fine-tuned from mistralai/Mistral-7B-v0.3. The card does not expose the exact served config or tensor layout needed for a bounds profile." }, { "label": "AllenAI WildGuard gated config and tensor-index access checks", "url": "https://huggingface.co/allenai/wildguard/raw/cbba4823f3e8020e5a74a5e29bf85072def6f2ff/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Anonymous raw requests for README.md, config.json, and model.safetensors.index.json returned 401. With the configured HF CLI identity, hf download README.md succeeded, but hf download config.json and model.safetensors.index.json returned: Access denied. This repository requires approval. The public API exposes only MistralForCausalLM/model_type mistral and tokenizer metadata, not layer count, KV heads, context length, tied embeddings, or tensor grouping." }, { "label": "Mistral 7B v0.3 lineage reference", "url": "https://huggingface.co/mistralai/Mistral-7B-v0.3/raw/caa1feb0e54d415e2df31207e5f4e273e33509b1/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The public Mistral-7B-v0.3 config records a 32-layer MistralForCausalLM BF16 base with 32 attention heads, 8 KV heads, 4096 hidden size, 32768 max positions, and untied embeddings. This is only lineage context; the WildGuard profile remains unsupported because its own gated config and tensor headers were not directly verified." } ], "unsupported_reason": "The repo's config and tensor headers are gated and inaccessible to the configured audit identity, so KV geometry, max context, tied embeddings, vocabulary changes, and swept ordinary text-decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct WildGuard config and safetensors header evidence is available." }, { "id": "andycurrent--gemma-3-1b-it-glm-4-7-flash-heretic-uncensored-thinking-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "title": "Gemma 3 1B GLM 4.7 Flash Heretic GGUF F16", "summary": "Audited memory-side bounds profile for the F16 GGUF artifact in the Andycurrent Gemma 3 1B derivative repo.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-1b-it", "relation": "derived_package", "source": "Hugging Face model card metadata and public GGUF header metadata", "config_compatible": false, "notes": "The repo card records google/gemma-3-1b-it as the base model, but the F16 GGUF metadata records DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated as an intermediate base and TeichAI/glm-4.7-2000x as a dataset. The gated Google base config was not accessible in this audit environment, so this profile uses the public GGUF header as the architecture source instead of copying a base config." }, "architecture": { "canonical_architecture_id": "gemma-3-1b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.999885952, "swept_params_b": 0.999885952, "resident_weight_gb": 2.006574688, "swept_weight_gb": 2.000040448, "auxiliary_resident_weight_gb": 0.00653424, "resident_parameter_scope": "F16 GGUF linked file size for the selected artifact", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans; token_embd.weight is also the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, header, and alignment padding are resident in the artifact file but not swept as model tensors", "notes": "The profile targets the F16 GGUF file. Header tensor spans total 2.000040448 GB, while the linked file size is 2.006574688 GB. The absence of output.weight means token_embd.weight remains in swept decode traffic as the tied output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 4, "kv_heads": 1, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 26 blocks, that yields four full-context global layers." }, { "kind": "sliding_window", "layers": 22, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The F16 GGUF metadata records gemma3.attention.sliding_window 512 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head and window metadata." }, "notes": "Dense Gemma3 GGUF profile audited from the selected F16 artifact, not from the inaccessible gated Google config." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains several GGUF quantizations. This profile intentionally targets Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_F16.gguf because the scrape and HF GGUF metadata point at the F16-size artifact and the card does not identify a smaller preferred default." }, "evidence": [ { "label": "Andycurrent Gemma 3 1B GGUF API metadata", "url": "https://huggingface.co/api/models/Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "max_context_tokens" ], "notes": "At commit 38eaddc2791eb42fafd4fd54205653c6cde70c86, the API records a Gemma-license text-generation GGUF repo with base_model google/gemma-3-1b-it. The API GGUF block reports architecture gemma3, 999885952 parameters, 32768 context length, and totalFileSize 2006574688 for the selected GGUF metadata." }, { "label": "Andycurrent Gemma 3 1B GGUF model card", "url": "https://huggingface.co/Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "weight_format" ], "notes": "The card describes a 1B instruction-tuned derivative of google/gemma-3-1b-it, says the model inherits the base context window, and states that quantized versions such as GGUF may be used depending on deployment stack. It does not recommend one specific GGUF quantization as the default." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 26 blocks recorded by the GGUF header." }, { "label": "Gemma 3 1B IT gated base config access check", "url": "https://huggingface.co/google/gemma-3-1b-it/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Unauthenticated raw config returned an access-denied response, and hf download google/gemma-3-1b-it config.json with the configured CLI token returned access denied because the repository requires approval. The profile therefore does not claim a direct base-config comparison." }, { "label": "Gemma 3 1B GLM 4.7 F16 GGUF range-read tensor index", "url": "https://huggingface.co/Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF/resolve/38eaddc2791eb42fafd4fd54205653c6cde70c86/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format" ], "notes": "HF linked-object metadata reports 2.006574688 GB for the F16 GGUF file. A 16MB range-read of the GGUF v3 header found 48 metadata entries and 340 tensors. Tensor spans sum to 2.000040448 GB: 1.99950336 GB F16 tensors and 0.000537088 GB F32 tensors. Metadata/header/alignment padding accounts for 0.00653424 GB resident-only bytes. The metadata records gemma3.block_count 26, context_length 32768, embedding_length 1152, feed_forward_length 6912, attention.head_count 4, attention.head_count_kv 1, attention.key_length 256, attention.value_length 256, and attention.sliding_window 512. token_embd.weight has shape [1152, 262144] and no output.weight tensor is stored." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, repo card text, gated-base access checks, Google Gemma 3 architecture documentation, HF linked-object metadata, and direct GGUF header/tensor-index range reads." }, "notes": "This profile corrects the scraped row's 2048-token context estimate to the 32768-token context recorded in the GGUF metadata and uses exact F16 GGUF artifact bytes instead of an inferred parameter-width estimate." }, { "id": "antirez--deepseek-v4-gguf", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "antirez/deepseek-v4-gguf", "title": "DeepSeek V4 Flash GGUF Q2 DS4 Imatrix", "summary": "Audited memory-side bounds profile for the preferred DS4 q2-imatrix GGUF package of DeepSeek V4 Flash.", "model_family": "deepseek-v4-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V4-Flash", "relation": "quantized", "source": "Hugging Face model card base_model metadata and DS4 GGUF model card", "config_compatible": true, "notes": "The profile embeds the DeepSeek V4 Flash architecture and the preferred DS4 q2-imatrix serving artifact." }, "architecture": { "canonical_architecture_id": "deepseek-v4-flash", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 86.720111488, "main_resident_weight_gb": 85.655715904, "auxiliary_resident_weight_gb": 1.064395584, "fixed_weight_gb": 7.7423248, "routed_expert_weight_gb": 0.304349184, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "gguf_q2_imatrix_file_size_and_tensor_index", "traffic_scope": "ordinary DS4 decode through q2-imatrix GGUF tensors, excluding the resident-only input embedding and GGUF metadata", "auxiliary_scope": "token_embd.weight and GGUF metadata/padding are resident for the artifact but not swept for each ordinary decode token", "shared_expert_notes": "The DeepSeek config records one shared expert. Shared expert tensors are stored as Q8_0 and included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "The original bounds note used rounded 284B total / 13B active Q2-style parameters. This production profile uses exact linked file size and range-read GGUF tensor-index spans for the preferred q2-imatrix artifact." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00314, "notes": "V1 compressed-state coefficient from the original bounds note worked example." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00121875, "notes": "V1 compressed-state read coefficient chosen to match the DS4 Q2 worked-example ceiling." }, "notes": "DeepSeek V4 uses compressed/sparse long-context attention. Bounds Engine v1 uses the audited coefficients from the source note rather than a generic compression ratio." }, "notes": "This profile targets the preferred DS4 q2-imatrix GGUF artifact, not a rounded parameter-count approximation." }, "serving": { "weight_format": "q2_mixed", "weight_bytes_per_param": 0.25, "kv_store_format": "ds4_compressed", "kv_store_bytes_per_scalar": 1, "kv_read_format": "ds4_compressed", "kv_read_bytes_per_scalar": 1, "runtime_format": "ds4-gguf-q2-imatrix-memory-bound", "dequantization_notes": "The memory-side bound charges exact q2-imatrix GGUF tensor spans for decode traffic and linked artifact file size for residency. DS4 kernel, dequantization, graph, and scheduler overheads are outside Bounds Engine v1.", "notes": "The preferred DS4 q2-imatrix package uses mixed IQ2_XXS, Q2_K, Q8_0, F16, F32, and I32 tensor classes. weight_bytes_per_param records the nominal q2 payload width; the audited adapter uses exact GGUF byte spans." }, "evidence": [ { "label": "DeepSeek V4 Flash model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "max_context_tokens", "compressed_attention" ], "notes": "The downloads table states 284B total, 13B activated, 1M context, and mixed FP4/FP8 precision for DeepSeek V4 Flash." }, { "label": "DeepSeek V4 Flash config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/main/config.json", "source_type": "config", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention" ], "notes": "The config records 256 routed experts, 6 experts per token, 1 shared expert, 1048576 max position embeddings, and compression-ratio metadata." }, { "label": "DS4 GGUF model card", "url": "https://huggingface.co/antirez/deepseek-v4-gguf", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "weight_format", "runtime_format" ], "notes": "The card identifies the q2 and q2-imatrix files as 128GB-machine packages and documents the mixed IQ2_XXS, Q2_K, Q8_0, F16, F32, and I32 tensor classes." }, { "label": "DS4 engine README", "url": "https://github.com/antirez/ds4", "source_type": "vendor_doc", "supports": [ "runtime_format", "preferred_artifact", "resident_weight_gb" ], "notes": "The engine README says to prefer q2-imatrix for 96/128GB machines and identifies this repo as the source for the DS4-specific GGUFs." }, { "label": "DS4 q2-imatrix GGUF range-read tensor index", "url": "https://huggingface.co/antirez/deepseek-v4-gguf/resolve/main/DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "HF linked-object metadata reports 86.720111488 GB for the q2-imatrix file. A 16MB range-read of the GGUF v3 header found 1328 tensors and 62 metadata entries. Tensor spans sum to 86.714777664 GB; routed expert tensors sum to 77.913391104 GB across 43 layers and 256 uniform experts, or 0.304349184 GB per expert index. Non-embedding fixed decode tensors sum to 7.7423248 GB. token_embd.weight plus GGUF metadata/padding accounts for 1.064395584 GB of resident-only bytes." }, { "label": "Original local frontier bounds note", "url": "https://github.com/osolmaz/onurclaw/blob/main/docs/2026-06-30-local-frontier-model-bounds.md", "source_type": "manual_review", "supports": [ "worked_example_parameters", "bounds_regression_target" ], "notes": "This evidence preserves the original rounded worked-example target, but the production profile intentionally uses exact q2-imatrix artifact bytes instead of the rounded 284B/13B Q2 approximation." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from DeepSeek config/card data, DS4 documentation, HF linked-object metadata, a range-read GGUF tensor index, and original bounds-note context." }, "notes": "The base DeepSeek repo reports FP4/FP8 mixed precision. This profile is only for the preferred antirez DS4 q2-imatrix GGUF package." }, { "id": "apolo13x--qwen3-5-9b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "apolo13x/Qwen3.5-9B-NVFP4", "title": "Apolo13x Qwen3.5 9B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the Apolo13x llm-compressor NVFP4 package of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model metadata, llm-compressor config/recipe, served config comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The repo card and API identify Qwen/Qwen3.5-9B as the quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds llm-compressor NVFP4 metadata while preserving the base hybrid attention geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 12.36044208, "swept_weight_gb": 8.927602432, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct llm-compressor NVFP4 safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from the two safetensors shard headers because this llm-compressor package mixes packed U8 NVFP4 MLP tensors, F8_E4M3 scale tensors, tiny F32 scalar scale tensors, and unquantized BF16 embeddings, lm_head, vision, MTP, linear-attention, self-attention, and norms. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 9.653104368B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The llm-compressor artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, dequantization, and NVFP4 activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The config records kv_cache_scheme null, so KV cache is charged at BF16." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 1.280462906935391, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llm-compressor-nvfp4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored llm-compressor NVFP4 safetensors bytes: packed U8 MLP weights, F8_E4M3 scales, F32 scalar scales, and unquantized BF16 tensors from the safetensors headers. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors NVFP4 pack quantization with group size 16 and kv_cache_scheme null. The recipe ignores lm_head, visual modules, MTP modules, linear-attention modules, and self-attention modules, so only MLP Linear modules are quantized." }, "evidence": [ { "label": "Apolo13x Qwen3.5 9B NVFP4 API metadata", "url": "https://huggingface.co/api/models/apolo13x/Qwen3.5-9B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 83fa19092744f3819c129ad41d21190cfa67bae1, the live API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B, with quantized, nvfp4, fp4, 4-bit, vLLM, llm-compressor, endpoints_compatible, compressed-tensors, and region:us tags. Current downloads are 334541. The API safetensors block reports F32: 192, BF16: 4821266160, F8_E4M3: 301989888, U8: 2415919104, and total: 7539175344 storage-accounting elements." }, { "label": "Apolo13x Qwen3.5 9B NVFP4 model card", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "quantized_module_scope" ], "notes": "The card states this is a quantized version of Qwen/Qwen3.5-9B that accepts text and images, quantized with llm-compressor using 512 calibration samples from nvidia/Nemotron-Post-Training-Dataset-v2. It reports NVFP4, max calibration sequence length 4096, vLLM 0.17.0 support, and about 97.3% average accuracy recovery." }, { "label": "Apolo13x Qwen3.5 9B NVFP4 config", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors NVFP4 pack quantization with 4-bit float weights/activations and group_size 16, text_config model_type qwen3_5_text, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, one MTP layer, and kv_cache_scheme null." }, { "label": "Apolo13x Qwen3.5 9B NVFP4 recipe", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/recipe.yaml", "source_type": "config", "supports": [ "serving", "quantization", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier targets [Linear], scheme NVFP4, and ignores lm_head, visual modules, MTP modules, in_proj_a/b/qkv/z, linear_attn.out_proj, and self_attn modules. This leaves the language MLP Linear modules as the quantized scope." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison found matching top-level identity fields and text geometry fields between the base BF16 repo and this llm-compressor NVFP4 artifact. Both configs record the same hybrid full-attention/linear-attention layer pattern and one MTP layer. The only checked text-field difference is that the Apolo artifact explicitly records tie_word_embeddings false inside text_config." }, { "label": "Apolo13x Qwen3.5 9B NVFP4 safetensors headers", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 12360442080 bytes across two shards. Direct range reads of both shard headers found 1063 tensors totaling exactly 12.360442080 GB: BF16 9.642532320 GB, U8 2.415919104 GB, F8_E4M3 0.301989888 GB, and F32 0.000000768 GB. Linked-object HEAD checks resolved both shards to 12.360573952 GB, leaving 131872 bytes of safetensors header/container overhead outside tensor payloads. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 8.927602432 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.432839648 GB. Header buckets are language linear-attention tensors 3.235390464 GB, language MLP tensors 2.717909760 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, language self-attention tensors 0.939532288 GB, visual 0.912020960 GB, MTP 0.486581248 GB, language layer/norm tensors 0.000532480 GB, and final language norm 0.000008192 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors NVFP4 config and recipe, current base config comparison, direct range-read safetensors headers, linked-object HEAD checks, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted the unquantized embeddings, output head, linear-attention, self-attention, MTP, visual tensors, and selected language tensors." }, { "id": "axionml--qwen3-5-9b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "AxionML/Qwen3.5-9B-NVFP4", "title": "AxionML Qwen3.5 9B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the AxionML ModelOpt NVFP4 package of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, ModelOpt config, and served config comparison", "config_compatible": true, "notes": "The artifact records Qwen/Qwen3.5-9B as its quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds ModelOpt NVFP4 quantization metadata while preserving the base hybrid attention geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 9.360860576, "swept_weight_gb": 5.928020928, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct ModelOpt NVFP4 safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from the single safetensors header because the ModelOpt package mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, tiny F32 scalar scale tensors, and unquantized BF16 embeddings, lm_head, vision, MTP, conv1d exclusions, and norms. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 9.653104368B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The ModelOpt artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, dequantization, and NVFP4 activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The model card and hf_quant_config record no KV cache quantization, so KV cache is charged at BF16." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.9697254084428232, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "modelopt-nvfp4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4 safetensors bytes: packed U8 weights, F8_E4M3 scales, F32 scalar scales, and unquantized BF16 tensors from the safetensors header. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config and hf_quant_config record NVFP4 weight and activation quantization for Linear modules with group size 16, while excluding lm_head, visual modules, MTP modules, and linear-attention conv1d modules. KV cache quantization is null." }, "evidence": [ { "label": "AxionML Qwen3.5 9B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/AxionML/Qwen3.5-9B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 97aef92393f126bf649f310cd40861be8dad3279, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B, with AxionML, ModelOpt, NVFP4, sglang, endpoints_compatible, and region:us tags. Current downloads are 748434. The API safetensors block reports BF16: 2734599920, F8_E4M3: 432406528, U8: 3459252224, and total: 6626258672 storage-accounting elements." }, { "label": "AxionML Qwen3.5 9B NVFP4 config", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/raw/97aef92393f126bf649f310cd40861be8dad3279/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, ModelOpt NVFP4 quantization with 4-bit float weights/activations and group_size 16, text_config model_type qwen3_5_text, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "AxionML Qwen3.5 9B NVFP4 hf_quant_config", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/raw/97aef92393f126bf649f310cd40861be8dad3279/hf_quant_config.json", "source_type": "config", "supports": [ "serving", "quantization" ], "notes": "The ModelOpt quantization sidecar records quant_algo NVFP4, kv_cache_quant_algo null, group_size 16, and excludes lm_head, visual modules, MTP modules, and every linear_attn.conv1d module from NVFP4 quantization." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching top-level identity fields and text geometry fields between the base BF16 repo and this ModelOpt NVFP4 artifact. Both configs record the same hybrid full-attention/linear-attention layer pattern and one MTP layer." }, { "label": "AxionML Qwen3.5 9B NVFP4 safetensors header", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/resolve/97aef92393f126bf649f310cd40861be8dad3279/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "A direct range-read of the single safetensors header found 1520 tensors. Stored tensor bytes sum to 9.360860576 GB: BF16 5.469199840 GB, U8 3.459252224 GB, F8_E4M3 0.432406528 GB, and F32 0.000001984 GB. The file content length is 9.361048680 GB, with 188104 bytes of safetensors header/container overhead outside tensor payloads. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 5.928020928 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.432839648 GB. Header buckets are language MLP 2.717909760 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, language linear-attention tensors 0.911091648 GB, visual 0.912020960 GB, MTP 0.486581248 GB, self-attention tensors 0.264249600 GB, and language layer/norm tensors 0.000532480 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned ModelOpt NVFP4 config and sidecar, current base config comparison, direct range-read safetensors header, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted ModelOpt scales plus unquantized embeddings, output head, conv1d exclusions, MTP, visual, and selected language tensors." }, { "id": "baidu--ernie-4-5-vl-28b-a3b-pt", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "baidu/ERNIE-4.5-VL-28B-A3B-PT", "title": "Baidu ERNIE 4.5 VL 28B A3B PT", "summary": "Audited memory-side ordinary text-decode bounds profile for the PyTorch ERNIE 4.5 VL 28B A3B MoE repo.", "model_family": "ernie4.5-vl-moe", "base_model_proof": { "base_model": "baidu/ERNIE-4.5-VL-28B-A3B-PT", "relation": "base", "source": "Hugging Face API metadata, model card, served config, custom runtime code, and direct safetensors header grouping", "config_compatible": true, "notes": "This profile targets the published PyTorch ERNIE 4.5 VL 28B A3B repo directly. The served config, custom Ernie4_5_VLMoeForConditionalGeneration implementation, and safetensors headers are used as the authoritative architecture and byte sources." }, "architecture": { "canonical_architecture_id": "ernie-4-5-vl-28b-a3b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 58.816290304, "main_resident_weight_gb": 43.659733504, "auxiliary_resident_weight_gb": 15.1565568, "fixed_weight_gb": 2.891098624, "routed_expert_weight_gb": 0.63700992, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f16_f32", "traffic_scope": "ordinary text decode through the tied output embedding, model.layers, model.norm, text gate, shared experts, and six selected text experts; vision encoder, resampler, vision expert pool, and vision gate are resident-only for this workload", "auxiliary_scope": "vision_model, model.resampler_model, the second 64-expert vision pool, and the second multimodal gate are resident in the VLM package but are not swept for ordinary text-token decode after any visual prefill", "shared_expert_notes": "The config records moe_num_shared_experts 2. Shared expert tensors are outside the routed experts namespace and run on every text token, so they are included in fixed_weight_gb.", "notes": "Header-derived stored bytes are used because the checkpoint mixes BF16 language weights, F16 vision weights, and F32 router/stat tensors. The config records two 64-expert pools, one text and one vision. The text-only decode path calls mlp_text with is_multimodel=false, so routing selects from the first 64 text experts only. The tied model.embed_tokens.weight is charged in fixed traffic because the config sets tie_word_embeddings true and the checkpoint has no separate lm_head.weight." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 28 text layers, 20 attention heads, 4 KV heads, hidden size 2560, and no sliding-window or recurrent text-state cache. Bounds Engine v1 charges full-context BF16 K and V streams for every text layer." }, "notes": "Ernie4_5_VLMoeForConditionalGeneration is multimodal. This profile models ordinary generated text-token decode after any image/video prefill, not vision encoder or visual-token prefill throughput." }, "serving": { "weight_format": "mixed_bf16_f16_f32", "weight_bytes_per_param": 2.0006021107785, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-ernie-vl-moe-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges stored BF16/F16/F32 safetensors bytes; router compute, expert compute, vision encoder execution, resampler prefill, activation traffic, and cache writes are outside this memory-side bound.", "notes": "The model card recommends Transformers with trust_remote_code and vLLM >= 0.11.2. The config records torch_dtype bfloat16 and no KV-cache quantization scheme, so KV is charged as BF16." }, "evidence": [ { "label": "ERNIE 4.5 VL 28B A3B PT API metadata", "url": "https://huggingface.co/api/models/baidu/ERNIE-4.5-VL-28B-A3B-PT", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "dtype_split", "serving" ], "notes": "At commit e3815e65c607ea211bfe21b46ab0cd264b76731c, the live API records a public non-gated Apache-2.0 Transformers image-text-to-text repo with ERNIE4.5, custom_code, endpoints_compatible, region:us, and 215890 downloads. The API safetensors summary records F32 8850816, BF16 28760010240, F16 630433280, and total 29399294336." }, { "label": "ERNIE 4.5 VL 28B A3B PT model card", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/README.md", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned card describes ERNIE-4.5-VL-28B-A3B as a multimodal MoE model with 28B total parameters, 3B activated parameters, 28 layers, 20 Q heads, 4 KV heads, 64 text experts with 6 activated, 64 vision experts with 6 activated, 2 shared experts, 131072 context length, PyTorch '-PT' weights, Transformers usage with trust_remote_code, and vLLM serving with vLLM >= 0.11.2." }, { "label": "ERNIE 4.5 VL 28B A3B PT config", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Ernie4_5_VLMoeForConditionalGeneration, ernie4_5_moe_vl, bfloat16 runtime dtype, tie_word_embeddings true, hidden size 2560, 28 layers, 20 attention heads, 4 KV heads, 128 head dimension, max_position_embeddings 131072, moe_num_experts [64, 64], moe_intermediate_size [1536, 512], moe_k 6, moe_num_shared_experts 2, moe_layer_start_index [1, 1], moe_layer_end_index [29, 28], and a 32-layer 1280-wide vision_config." }, { "label": "ERNIE 4.5 VL custom runtime review", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/modeling_ernie4_5_vl.py", "source_type": "manual_review", "supports": [ "traffic_scope", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "tied_output_projection" ], "notes": "Manual review found layer 0 uses dense Ernie4_5_MLP, layers 1-27 use MOELayer with shared_experts, and the text-only forward path calls self.mlp_text()(hidden_states, None, is_multimodel=false). In that mode TopKGate.get_gate_weight returns weight for the first 64-expert text pool. The VLM class replaces lm_head with a Linear and post_init ties weights; the checkpoint stores no separate lm_head.weight, so model.embed_tokens.weight is the tied output projection." }, { "label": "ERNIE 4.5 VL 28B A3B PT safetensors index and shard headers", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 58.816290304 GB across 12 shards. Direct range-read safetensors headers found 11105 tensors totaling the same 58.816290304 GB: F32 0.035403264 GB, BF16 57.520020480 GB, and F16 1.260866560 GB. Ordinary text resident tensors, defined as tied model.embed_tokens.weight, model.layers, and model.norm excluding the resident-only vision expert pool and vision gate, sum to 43.659733504 GB. Auxiliary resident tensors, defined as vision_model, model.resampler_model, the second 64-expert vision pool, and the second multimodal gate, sum to 15.156556800 GB. Routed text expert tensors sum to 40.768634880 GB and divide exactly into 64 uniform text expert indexes of 0.637009920 GB. Fixed text-decode traffic including the tied output embedding, attention, dense layer 0 MLP, norms, text gates, shared experts, and MoE statics sums to 2.891098624 GB." } ], "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-06", "notes": "Manual one-model audit from pinned HF API/config/modeling/index evidence and direct safetensors header grouping." }, "notes": "This self-contained profile supersedes the generated metadata estimate. It uses exact stored tensor bytes and charges tied output embedding traffic, while keeping vision encoder and vision expert pool bytes resident-only for ordinary text decode." }, { "id": "bartowski--gemma-2-2b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/gemma-2-2b-it-GGUF", "title": "bartowski Gemma 2 2B IT GGUF F32", "summary": "Audited memory-side text-decode bounds profile for the API-selected F32 GGUF artifact of Gemma 2 2B IT.", "model_family": "gemma2-dense-gguf", "base_model_proof": { "base_model": "google/gemma-2-2b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, gated Google base-profile checks, selected GGUF header metadata, and the Transformers Gemma 2 implementation", "config_compatible": false, "notes": "The GGUF repo metadata identifies google/gemma-2-2b-it as the quantized base. The Google base profile is gated in this audit environment, so this profile uses the selected public GGUF header and pinned Gemma 2 implementation defaults as direct architecture evidence instead of copying the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-2-2b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.614341888, "swept_params_b": 2.614341888, "auxiliary_resident_params_b": 0, "resident_weight_gb": 10.463413856, "swept_weight_gb": 10.457367552, "auxiliary_resident_weight_gb": 0.006046304, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-2-2b-it-f32.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected F32 linked file is 10.463413856 GB. Header tensor spans total 10.457367552 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.006046304 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 13, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The pinned Transformers Gemma 2 config alternates sliding_attention on zero-indexed even layers and full_attention on zero-indexed odd layers. With 26 layers, this gives 13 full-context attention layers." }, { "kind": "sliding_window", "layers": 13, "kv_heads": 4, "head_dim": 256, "window_tokens": 4096, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records gemma2.attention.sliding_window 4096, and the pinned Gemma 2 attention path applies that window only to sliding_attention layers." } ], "notes": "Hybrid Gemma 2 text attention is represented as explicit full-context plus sliding-window KV components. The selected GGUF artifact does not declare quantized KV cache, so Bounds Engine v1 charges llama.cpp-style FP16 K/V cache storage and read traffic." }, "notes": "This profile models ordinary text decode for the API-selected F32 GGUF artifact. It uses public GGUF metadata and pinned Transformers Gemma 2 implementation evidence because the Google base config is gated." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4.002312744185354, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F32 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many smaller GGUF quantizations, including recommended Q/I-quant files. This profile intentionally targets gemma-2-2b-it-f32.gguf because the HF API gguf.totalFileSize exactly matches that linked object, following the selected-artifact rule used by the existing Bartowski GGUF profiles." }, "evidence": [ { "label": "bartowski Gemma 2 2B IT GGUF API metadata", "url": "https://huggingface.co/api/models/bartowski/gemma-2-2b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 855f67caed130e1befc571b52bd181be2e858883, the API records a public non-gated GGUF repo with base_model google/gemma-2-2b-it, Gemma license, region:us, 258944 downloads, GGUF architecture gemma2, 8192 context length, gguf.total 2614341888, and gguf.totalFileSize 10463413856." }, { "label": "bartowski Gemma 2 2B IT GGUF model card", "url": "https://huggingface.co/bartowski/gemma-2-2b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "quantization_recipe" ], "notes": "The card records this as a llama.cpp b3496 imatrix quantization repo by bartowski, based on google/gemma-2-2b-it. It lists many single-file GGUF variants; the F32 row is described as full F32 weights, while smaller Q/I-quant rows are separate artifacts." }, { "label": "Google Gemma 2 2B IT gated base profile", "url": "https://huggingface.co/google/gemma-2-2b-it", "source_type": "manual_review", "supports": [ "base_model_proof", "unsupported_gated_base" ], "notes": "The existing local unsupported profile for google/gemma-2-2b-it records API safetensors total 2614341888 parameters at commit 299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8, but config and tensor headers are gated. This GGUF derivative is audited from its public selected artifact instead of inferring from the gated base profile." }, { "label": "bartowski Gemma 2 2B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/bartowski/gemma-2-2b-it-GGUF/tree/855f67caed130e1befc571b52bd181be2e858883", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found F32 10463413856 bytes, Q8_0 2784495456 bytes, Q6_K_L 2294241120 bytes, Q6_K 2151393120 bytes, Q5_K_M 1923278688 bytes, Q5_K_S 1882543968 bytes, Q4_K_M 1708582752 bytes, Q4_K_S 1638651744 bytes, IQ4_XS 1566250848 bytes, Q3_K_L 1550436192 bytes, and IQ3_M 1393561440 bytes. The selected F32 artifact exactly matches API gguf.totalFileSize." }, { "label": "bartowski Gemma 2 2B IT F32 GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/gemma-2-2b-it-GGUF/resolve/855f67caed130e1befc571b52bd181be2e858883/gemma-2-2b-it-f32.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 288 tensors. The linked file is 10.463413856 GB. Tensor spans sum to 10.457367552 GB: token_embd.weight 2.359296 GB, blk.* tensors 8.098062336 GB, and output_norm.weight 0.000009216 GB. Metadata/tokenizer/header/file overhead accounts for 0.006046304 GB. Stored tensor bytes are entirely F32. The header records gemma2.block_count 26, context_length 8192, embedding_length 2304, feed_forward_length 9216, attention.head_count 8, attention.head_count_kv 4, attention key/value length 256, attention.sliding_window 4096, and no separate output.weight tensor." }, { "label": "Transformers Gemma 2 config and attention implementation", "url": "https://github.com/huggingface/transformers/blob/b70d02fc724d04c916832ca4ead03ff05e8fb1ee/src/transformers/models/gemma2/configuration_gemma2.py", "source_type": "config", "supports": [ "layer_pattern", "kv_adapter", "sliding_window" ], "notes": "The pinned Gemma2Config defaults match the GGUF geometry: 26 layers, 8 attention heads, 4 KV heads, 256 head dimension, 8192 max positions, and 4096 sliding window. Its __post_init__ creates sliding_attention for zero-indexed even layers and full_attention for zero-indexed odd layers. The pinned Gemma2Attention implementation sets self.sliding_window only for sliding_attention layers and passes that window to the attention kernel." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, existing gated-base profile evidence, HEAD checks for all GGUF linked file sizes, a direct GGUF header/tensor-index range read of the selected F32 artifact, and pinned Transformers Gemma 2 config/attention implementation review." }, "notes": "Use this profile for the API-selected Gemma 2 2B IT F32 GGUF artifact. Do not silently substitute the smaller Q4_K_M or other quantized artifacts; those require separate profiles with their own selected artifact bytes." }, { "id": "bartowski--google-gemma-4-26b-a4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/google_gemma-4-26B-A4B-it-GGUF", "title": "Bartowski Gemma 4 26B A4B IT GGUF IQ2_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected Bartowski IQ2_M GGUF artifact of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face API/model-card metadata, Google base profile/config geometry, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo metadata identifies this package as a GGUF derivative of google/gemma-4-26B-A4B-it. The selected IQ2_M GGUF header records the same Gemma 4 26B A4B text architecture: 30 layers, five full-attention layers, 25 sliding-attention layers, 128 experts, 8 routed experts per token, 1024-token sliding window, and 262144-token max context." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 10.700620704, "main_resident_weight_gb": 10.6847984, "auxiliary_resident_weight_gb": 0.015822304, "fixed_weight_gb": 1.470290368, "routed_expert_weight_gb": 0.071988344, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for google_gemma-4-26B-A4B-it-IQ2_M.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors; imatrix, mmproj, and MTP GGUF sidecars are not included unless explicitly loaded for another workload", "shared_expert_notes": "The GGUF header records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects google_gemma-4-26B-A4B-it-IQ2_M.gguf. A GGUF v3 range-read found 658 tensors and 54 metadata entries. Tensor spans total 10.684798400 GB, while the linked file is 10.700620704 GB. Routed expert tensors total 9.214508032 GB across 30 layers and 128 expert indexes, or 0.071988344 GB per expert index. Non-expert tensor spans total 1.470290368 GB, including token_embd.weight because the selected file has no separate output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. Gemma 4 full-attention layers use K=V behavior, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile targets the API-selected main IQ2_M text GGUF artifact. The imatrix file, multimodal projector files, speculative MTP files, and other quantized GGUF siblings are separate artifacts and should get separate workload profiles if loaded." }, "serving": { "weight_format": "q2_mixed", "weight_bytes_per_param": 0.42407008546509095, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq2-m-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, image handling, speculative MTP execution, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the IQ2_M GGUF because HF API gguf.totalFileSize exactly matches google_gemma-4-26B-A4B-it-IQ2_M.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Bartowski Gemma 4 26B A4B IT GGUF API metadata", "url": "https://huggingface.co/api/models/bartowski/google_gemma-4-26B-A4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit fabed3e586120477355eea23b92644540a79ce2f, the live API records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model google/gemma-4-26B-A4B-it, region:us, 106081 downloads, GGUF architecture gemma4, context_length 262144, gguf.total 25233142046, and gguf.totalFileSize 10700620704." }, { "label": "Bartowski Gemma 4 26B A4B IT GGUF model card", "url": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF/raw/fabed3e586120477355eea23b92644540a79ce2f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact" ], "notes": "The card metadata records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-26B-A4B-it, quantization by bartowski using llama.cpp, imatrix quantization, and GGUF sibling artifacts." }, { "label": "Google Gemma 4 26B A4B IT audited profile", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "kv_adapter" ], "notes": "The existing audited Google profile records 30 layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, attention_k_eq_v true, one shared expert, tied embeddings, and 262144 max positions." }, { "label": "Bartowski Gemma 4 26B A4B IT linked GGUF sizes", "url": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF/tree/fabed3e586120477355eea23b92644540a79ce2f", "source_type": "manual_review", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "Expanded tree and HEAD checks found google_gemma-4-26B-A4B-it-IQ2_M.gguf is 10.700620704 GB, exactly matching API gguf.totalFileSize. Other main siblings range from 9.656494368 GB to 26.859859360 GB; BF16 is split into a separate folder; imatrix, mmproj, and MTP files are separate artifacts." }, { "label": "Bartowski Gemma 4 26B A4B IT IQ2_M GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF/resolve/fabed3e586120477355eea23b92644540a79ce2f/google_gemma-4-26B-A4B-it-IQ2_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 658 tensors and 54 metadata entries. Tensor spans sum to 10.684798400 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015822304 GB. Tensor spans split into IQ2_S 4.819075712 GB, IQ4_NL 4.379139072 GB, Q5_K 0.785055744 GB, Q4_K 0.361912320 GB, IQ3_S 0.225737600 GB, Q6_K 0.061501440 GB, F32 0.046057408 GB, and Q8_0 0.006319104 GB. Non-expert tensor spans total 1.470290368 GB. Routed expert tensors total 9.214508032 GB across 30 layers and 128 expert indexes, or 0.071988344 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, K=V full-attention geometry, separate sliding-layer K/V projections, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from live HF API metadata, pinned model card metadata, the existing Google Gemma 4 base profile, selected linked-object HEAD checks, expanded linked-file metadata, and a direct GGUF header/tensor-index range read of the selected IQ2_M artifact." }, "notes": "Use this profile for the Bartowski API-selected IQ2_M GGUF text artifact in ordinary text-decode bounds. Do not infer BF16, Q4_K_M, Q8_0, imatrix, mmproj, or MTP sidecar footprints unless the workload profile explicitly selects and audits those files." }, { "id": "bartowski--llama-3-2-1b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/Llama-3.2-1B-Instruct-GGUF", "title": "bartowski Llama 3.2 1B Instruct GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Llama 3.2 1B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-1B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, and GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.2-1B-Instruct as the quantized base. The base raw config is gated in this audit environment, so this profile uses the selected public GGUF header as the direct architecture source instead of copying the base config." }, "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.235814432, "swept_params_b": 1.235814432, "resident_weight_gb": 2.47959536, "swept_weight_gb": 2.471764096, "auxiliary_resident_weight_gb": 0.007831264, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.2-1B-Instruct-f16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected F16 linked file is 2.479595360 GB. Header tensor spans total 2.471764096 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007831264 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 16-layer decoder with 8 KV heads and 64-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. It uses the public GGUF metadata for architecture because the Meta base config is gated." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many GGUF quantizations. This profile intentionally targets Llama-3.2-1B-Instruct-f16.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "bartowski Llama 3.2 1B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-1B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 067b946cf014b7c697f3654f621d577a3e3afd1c, the API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.2-1B-Instruct, Llama 3.2 license, region:us, 416350 downloads, GGUF architecture llama, 131072 context length, gguf.total 1235814432, and gguf.totalFileSize 2479595360." }, { "label": "bartowski Llama 3.2 1B Instruct GGUF model card", "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "quantization_recipe" ], "notes": "The card records this as a llama.cpp b3821 imatrix quantization repo by bartowski, based on meta-llama/Llama-3.2-1B-Instruct. It lists many single-file GGUF variants; the F16 row is described as full F16 weights." }, { "label": "Llama 3.2 1B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 9213176726f574b556790deb65791e0c5aa438b6, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 1235814400 parameters." }, { "label": "Llama 3.2 1B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "hf download of config.json with the configured CLI token returned: Access denied. This repository requires approval. The profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "bartowski Llama 3.2 1B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/tree/067b946cf014b7c697f3654f621d577a3e3afd1c", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found F16 2479595360 bytes, Q8_0 1321083008 bytes, Q6_K_L 1085415552 bytes, Q6_K 1021800576 bytes, Q5_K_L 975118464 bytes, Q5_K_M 911503488 bytes, Q5_K_S 892563584 bytes, Q4_K_L 871309440 bytes, Q4_K_M 807694464 bytes, Q3_K_XL 796139648 bytes, Q4_K_S 775647360 bytes, Q4_0 773025920 bytes, Q4_0 ARM variants 770928768 bytes, IQ4_XS 743141504 bytes, Q3_K_L 732524672 bytes, and IQ3_M 657289344 bytes. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "bartowski Llama 3.2 1B Instruct F16 GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/067b946cf014b7c697f3654f621d577a3e3afd1c/Llama-3.2-1B-Instruct-f16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 31 metadata entries and 147 tensors. The linked file is 2.479595360 GB. Tensor spans sum to 2.471764096 GB: token_embd.weight 0.525336576 GB, blk.* tensors 1.946419200 GB, output_norm.weight 0.000008192 GB, and rope_freqs.weight 0.000000128 GB. Metadata/tokenizer/header/file overhead accounts for 0.007831264 GB. Stored tensor bytes split into F16 2.471493632 GB and F32 0.000270464 GB. The header records llama.block_count 16, context_length 131072, embedding_length 2048, feed_forward_length 8192, attention.head_count 32, attention.head_count_kv 8, attention key/value length 64, rope.freq_base 500000, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, base-model API metadata, gated base-config access check, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Llama 3.2 1B Instruct F16 GGUF artifact. Do not infer the gated base config directly; the architecture evidence is the selected GGUF header metadata." }, { "id": "bartowski--llama-3-2-3b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/Llama-3.2-3B-Instruct-GGUF", "title": "bartowski Llama 3.2 3B Instruct GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Llama 3.2 3B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-3B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, and GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.2-3B-Instruct as the quantized base. The base raw config is gated in this audit environment, so this profile uses the selected public GGUF header as the direct architecture source instead of copying the base config." }, "architecture": { "canonical_architecture_id": "llama-3-2-3b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.212749888, "swept_params_b": 3.212749888, "resident_weight_gb": 6.43368784, "swept_weight_gb": 6.425850112, "auxiliary_resident_weight_gb": 0.007837728, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.2-3B-Instruct-f16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected F16 linked file is 6.433687840 GB. Header tensor spans total 6.425850112 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007837728 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 28-layer decoder with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. It uses the public GGUF metadata for architecture because the Meta base config is gated." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many GGUF quantizations. This profile intentionally targets Llama-3.2-3B-Instruct-f16.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "bartowski Llama 3.2 3B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 5ab33fa94d1d04e903623ae72c95d1696f09f9e8, the API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.2-3B-Instruct, Llama 3.2 license, region:us, 213102 downloads, GGUF architecture llama, 131072 context length, gguf.total 3212749888, and gguf.totalFileSize 6433687840." }, { "label": "bartowski Llama 3.2 3B Instruct GGUF model card", "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "quantization_recipe" ], "notes": "The card records this as a llama.cpp b3821 imatrix quantization repo by bartowski, based on meta-llama/Llama-3.2-3B-Instruct. It lists many single-file GGUF variants; the F16 row is described as full F16 weights." }, { "label": "Llama 3.2 3B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 0cb88a4f764b7a12671c53f0838cd831a0843b95, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 3212749824 parameters." }, { "label": "Llama 3.2 3B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/raw/0cb88a4f764b7a12671c53f0838cd831a0843b95/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Unauthenticated raw config access returned HTTP 401. The existing base-profile audit also records denied config and safetensors index access with the configured CLI token. This GGUF profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "bartowski Llama 3.2 3B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/tree/5ab33fa94d1d04e903623ae72c95d1696f09f9e8", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found F16 6433687840 bytes, Q8_0 3421899296 bytes, Q6_K_L 2739276320 bytes, Q6_K 2643853856 bytes, Q5_K_L 2417576480 bytes, Q5_K_M 2322154016 bytes, Q5_K_S 2269512224 bytes, Q4_K_L 2114800160 bytes, Q4_K_M 2019377696 bytes, Q4_K_S 1928200736 bytes, Q4_0 1921909280 bytes, Q4_0 ARM variants 1917190688 bytes, Q3_K_XL 1910770208 bytes, IQ4_XS 1829110304 bytes, Q3_K_L 1815347744 bytes, and IQ3_M 1599668768 bytes. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "bartowski Llama 3.2 3B Instruct F16 GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/5ab33fa94d1d04e903623ae72c95d1696f09f9e8/Llama-3.2-3B-Instruct-f16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 31 metadata entries and 255 tensors. The linked file is 6.433687840 GB. Tensor spans sum to 6.425850112 GB: token_embd.weight 0.788004864 GB, blk.* tensors 5.637832704 GB, output_norm.weight 0.000012288 GB, and rope_freqs.weight 0.000000256 GB. Metadata/tokenizer/header/file overhead accounts for 0.007837728 GB. Stored tensor bytes split into F16 6.425149440 GB and F32 0.000700672 GB. The header records llama.block_count 28, context_length 131072, embedding_length 3072, feed_forward_length 8192, attention.head_count 24, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 500000, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, base-model API metadata, gated base-config access check, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Llama 3.2 3B Instruct F16 GGUF artifact. Do not infer the gated base config directly; the architecture evidence is the selected GGUF header metadata." }, { "id": "bartowski--meta-llama-3-1-8b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "title": "Bartowski Meta Llama 3.1 8B Instruct GGUF F32", "summary": "Audited memory-side text-decode bounds profile for the API-selected F32 GGUF artifact of Meta Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, gated Meta base-profile checks, and selected GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Meta-Llama-3.1-8B-Instruct as the quantized base. The Meta base profile is gated in this audit environment, so this profile uses the selected public GGUF header as direct architecture evidence instead of copying the gated base config." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261312, "swept_params_b": 7.504924736, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 32.128885888, "swept_weight_gb": 30.019698944, "auxiliary_resident_weight_gb": 2.109186944, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Meta-Llama-3.1-8B-Instruct-f32.gguf", "swept_parameter_scope": "ordinary text decode excludes token_embd.weight input lookup and includes blk.* tensors, output.weight, output_norm.weight, and rope_freqs.weight from the selected F32 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file alignment are resident in the selected artifact but not swept as full matrices for each generated token", "notes": "The selected F32 linked file is 32.128885888 GB. Header tensor spans total 32.121045248 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007840640 GB. The main GGUF stores token_embd.weight separately from output.weight, so ordinary text decode excludes the input embedding lookup and charges the separate output projection." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 Llama blocks, 32 attention heads, 8 KV heads, 4096 hidden size, and 128 RoPE/head dimension. No sliding-window setting is present, so this profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F32 GGUF artifact. Smaller Q/I-quant artifacts in the repo require separate selected-artifact profiles." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4.000976386657341, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F32 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many smaller GGUF quantizations, including recommended Q/I-quant files. This profile intentionally targets Meta-Llama-3.1-8B-Instruct-f32.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "Bartowski Meta Llama 3.1 8B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit bf5b95e96dac0462e2a09145ec66cae9a3f12067, the API records a public non-gated text-generation GGUF repo with base_model meta-llama/Meta-Llama-3.1-8B-Instruct, llama3.1 license, region:us, 261959 downloads, GGUF architecture llama, 131072 context length, gguf.total 8030261312, and gguf.totalFileSize 32128885888." }, { "label": "Bartowski Meta Llama 3.1 8B Instruct GGUF model card", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/raw/bf5b95e96dac0462e2a09145ec66cae9a3f12067/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "quantization_recipe" ], "notes": "The card records this as a llama.cpp b3472 imatrix quantization repo by bartowski, based on meta-llama/Meta-Llama-3.1-8B-Instruct. Its artifact table describes the F32 file as full F32 weights and lists smaller Q/I-quant siblings as separate artifacts." }, { "label": "Meta Llama 3.1 8B Instruct gated base profile", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "unsupported_gated_base" ], "notes": "The existing local unsupported profile for meta-llama/Llama-3.1-8B-Instruct records API safetensors total 8030261248 parameters, but config and tensor headers are gated. This GGUF derivative is audited from its public selected artifact instead of inferring from the gated base profile." }, { "label": "Bartowski Meta Llama 3.1 8B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/bf5b95e96dac0462e2a09145ec66cae9a3f12067", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found F32 32128885888 bytes, Q8_0 8540775840 bytes, Q6_K_L 6850471328 bytes, Q6_K 6596011424 bytes, Q5_K_L 6057223584 bytes, Q5_K_M 5732992416 bytes, Q5_K_S 5599298976 bytes, Q4_K_L 5310637472 bytes, Q4_K_M 4920739232 bytes, Q3_K_XL 4781630880 bytes, Q4_K_S 4692673952 bytes, IQ4_NL 4677993888 bytes, Q4_0_* 4661216672 bytes, IQ4_XS 4447667616 bytes, Q3_K_L 4321961376 bytes, Q3_K_M 4018922912 bytes, IQ3_M 3784828320 bytes, Q2_K_L 3692160416 bytes, Q3_K_S 3664504224 bytes, IQ3_XS 3518752160 bytes, Q2_K 3179136416 bytes, and IQ2_M 2948285856 bytes. The selected F32 artifact exactly matches API gguf.totalFileSize." }, { "label": "Bartowski Meta Llama 3.1 8B Instruct F32 GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/bf5b95e96dac0462e2a09145ec66cae9a3f12067/Meta-Llama-3.1-8B-Instruct-f32.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 29 metadata entries and 292 tensors. The linked file is 32.128885888 GB. Tensor spans sum to 32.121045248 GB: token_embd.weight 2.101346304 GB, output.weight 2.101346304 GB, blk.* tensors 27.918336000 GB, output_norm.weight 0.000016384 GB, and rope_freqs.weight 0.000000256 GB. Metadata/tokenizer/header/file overhead accounts for 0.007840640 GB. Stored tensor bytes are entirely F32. The header records general.architecture llama, llama.block_count 32, context_length 131072, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, RoPE dimension/head dimension 128, RoPE base 500000, and a separate output.weight tensor." }, { "label": "Bartowski Meta Llama 3.1 8B Instruct GGUF missing repo config check", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/raw/bf5b95e96dac0462e2a09145ec66cae9a3f12067/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A pinned repo-local config.json request returned 404. The profile therefore uses the selected GGUF header directly and does not claim repo-local config evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, existing gated-base profile evidence, linked GGUF file sizes, pinned missing repo-config check, and direct selected F32 GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected Meta Llama 3.1 8B Instruct F32 GGUF artifact. Do not silently substitute the smaller Q4_K_M or other quantized artifacts; those require separate profiles with their own selected artifact bytes." }, { "id": "bartowski--qwen-qwen3-6-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bartowski/Qwen_Qwen3.6-35B-A3B-GGUF", "title": "Bartowski Qwen3.6 35B A3B GGUF IQ1_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected Bartowski IQ1_M GGUF artifact of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-35B-A3B. The selected IQ1_M GGUF header records the same audited Qwen3.6 text geometry as the Qwen config: 40 ordinary text blocks, every fourth ordinary layer using full attention, 256 routed experts, 8 routed experts per token, one always-on shared expert path, and one resident MTP draft block." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 9.421649536, "main_resident_weight_gb": 8.345831936, "auxiliary_resident_weight_gb": 1.0758176, "fixed_weight_gb": 1.205029376, "routed_expert_weight_gb": 0.02789376, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf linked file size including GGUF metadata, tokenizer, header, tensor spans, and resident MTP draft tensors", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, ordinary blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected IQ1_M GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.40 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; separate mmproj and MTP sidecars are not included unless explicitly loaded for another workload", "shared_expert_notes": "The Qwen model card states 8 routed plus 1 shared expert, and the GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected API artifact is the IQ1_M file, which the model card describes as extremely low quality and not recommended. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 40 ordinary layers with every fourth layer using full attention, giving 10 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ1_M GGUF artifact after any multimodal prefill, with speculative MTP disabled. The separate mmproj GGUF files and optional MTP speculative execution path require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.265359324315047, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq1-m-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Bartowski Qwen3.6 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 5c2410d71524f4f72b023ce8daf7a80528226d5f records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-35B-A3B, 120082 downloads, region:us, imatrix metadata, GGUF architecture qwen35moe, 262144 context length, gguf.total 35505251456, and gguf.totalFileSize 9421649536. The API totalFileSize matches Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf, so this profile targets that artifact." }, { "label": "Bartowski Qwen3.6 35B A3B GGUF model card", "url": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/raw/5c2410d71524f4f72b023ce8daf7a80528226d5f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope", "routed_experts", "shared_experts_per_token" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.6-35B-A3B, quantization by bartowski using llama.cpp b9222, imatrix quantization, and MTP support via --spec-type draft-mtp. Its quant table describes Q4_K_M as the default recommended size for most use cases, while IQ1_M is 9.42GB and explicitly not recommended; this profile still targets IQ1_M because that is the current HF API-selected artifact." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable base API commit records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, and one MTP layer." }, { "label": "Bartowski Qwen3.6 35B A3B linked-object HEAD checks", "url": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/tree/5c2410d71524f4f72b023ce8daf7a80528226d5f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of GGUF siblings found selected IQ1_M 9.421649536 GB, Q4_K_M 22.285080192 GB, Q8_0 37.812647552 GB, separate mmproj BF16 0.902822240 GB, MTP Q4_0 sidecar 1.190098816 GB, and MTP Q8_0 sidecar 1.989597056 GB." }, { "label": "Bartowski Qwen3.6 35B A3B IQ1_M GGUF range-read tensor index", "url": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/resolve/5c2410d71524f4f72b023ce8daf7a80528226d5f/Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 49 metadata entries and 753 tensors. The linked file is 9.421649536 GB. Tensor spans sum to 9.410658816 GB: output.weight plus output_norm.weight 0.349642752 GB, token_embd.weight 0.166871040 GB, ordinary non-routed blk.0-39 tensors 0.855386624 GB, routed expert tensors 7.140802560 GB, and blk.40 MTP tensors 0.897955840 GB. Metadata/tokenizer/header/file overhead accounts for 0.010990720 GB. Tensor spans split into IQ1_M 6.602358784 GB, Q8_0 0.963706880 GB, IQ2_XXS 0.622854144 GB, Q4_K 0.403439616 GB, Q5_K 0.349634560 GB, Q6_K 0.196116480 GB, Q2_K 0.166871040 GB, F32 0.104624640 GB, and BF16 0.001052672 GB. The GGUF metadata records qwen35moe.block_count 41, nextn_predict_layers 1, context_length 262144, attention.head_count 16, attention.head_count_kv 2, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, expert_count 256, expert_used_count 8, and expert_shared_feed_forward_length 512." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned Qwen base config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the API-selected IQ1_M artifact." }, "notes": "Use this profile for the Bartowski Qwen3.6 35B A3B IQ1_M main GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer the card-recommended Q4_K_M artifact, multimodal projector residency, sidecar MTP residency, or speculative MTP acceleration unless the workload profile explicitly selects those paths." }, { "id": "batiai--qwen3-6-27b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "batiai/Qwen3.6-27B-GGUF", "title": "BatiAI Qwen3.6 27B GGUF IQ3_XXS", "summary": "Audited memory-side text-decode bounds profile for the API-selected BatiAI IQ3_XXS GGUF artifact of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, pinned Qwen base config, selected linked-object size metadata, and selected GGUF header metadata", "config_compatible": true, "notes": "The BatiAI card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-27B. The selected IQ3_XXS GGUF header records the same Qwen3.6 text geometry as the Qwen config, with 64 ordinary text layers and no MTP, mmproj, vision, or draft tensors in the main file." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 11.1863704, "swept_weight_gb": 10.629072896, "auxiliary_resident_weight_gb": 0.557297504, "resident_parameter_scope": "selected Qwen-Qwen3.6-27B-IQ3_XXS.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected IQ3_XXS GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj sidecars are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen-Qwen3.6-27B-IQ3_XXS.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. The selected linked file is 11.186370400 GB. Header tensor spans total 11.175376896 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010993504 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, and ordinary blk.0-63 tensors, with no MTP, mmproj, vision, or draft tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ3_XXS GGUF artifact after any multimodal prefill. The separate mmproj-BF16/Q6_K GGUF files require a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.4159120701532175, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq3-xxs-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen-Qwen3.6-27B-IQ3_XXS.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "BatiAI Qwen3.6 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/batiai/Qwen3.6-27B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 8057b95492cfe844051386e1e4f04ee9c99f7b95 records a public Apache-2.0 text-generation GGUF repo with base_model Qwen/Qwen3.6-27B, 154057 downloads, region:us, imatrix metadata, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 11186370400. The API totalFileSize exactly matches Qwen-Qwen3.6-27B-IQ3_XXS.gguf, so this profile targets that artifact." }, { "label": "BatiAI Qwen3.6 27B GGUF model card", "url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF/raw/8057b95492cfe844051386e1e4f04ee9c99f7b95/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "weight_format", "selected_artifact", "auxiliary_scope" ], "notes": "The card records Apache-2.0 licensing, base_model Qwen/Qwen3.6-27B, Qwen3.6 text geometry, BatiAI imatrix GGUF quantization, Mac/Ollama tags, and separate mmproj sidecars for multimodal use. Quick-start examples include iq3 for tight 16GB Macs and iq4/q4 recommendations for 24GB Macs, but the HF API-selected artifact is IQ3_XXS." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP training settings." }, { "label": "BatiAI Qwen3.6 27B GGUF linked-file size metadata", "url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF/tree/8057b95492cfe844051386e1e4f04ee9c99f7b95", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "Expanded tree metadata records Qwen-Qwen3.6-27B-IQ3_XXS.gguf 11.186370400 GB, IQ4_XS 15.082506080 GB, Q3_K_M 13.301442080 GB, Q4_K_M 16.547399200 GB, Q6_K 22.082528800 GB, imatrix.dat 0.013642688 GB, mmproj-Qwen-Qwen3.6-27B-BF16.gguf 0.931146080 GB, and mmproj-Qwen-Qwen3.6-27B-Q6_K.gguf 0.618390752 GB. IQ3_XXS exactly matches API gguf.totalFileSize and the mmproj sidecars are not part of the selected main text artifact." }, { "label": "BatiAI Qwen3.6 27B IQ3_XXS GGUF range-read tensor index", "url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF/resolve/8057b95492cfe844051386e1e4f04ee9c99f7b95/Qwen-Qwen3.6-27B-IQ3_XXS.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 48 metadata entries and 851 tensors. The linked file is 11.186370400 GB. Tensor spans sum to 11.175376896 GB: output.weight plus output_norm.weight 0.874106880 GB, token_embd.weight 0.546304000 GB, and ordinary blk.0-63 tensors 9.754966016 GB. Metadata/tokenizer/header/file overhead accounts for 0.010993504 GB. Tensor spans split into IQ3_XXS 7.716065280 GB, Q4_K 1.462763520 GB, Q5_K 0.874086400 GB, IQ3_S 0.762572800 GB, IQ2_S 0.349306880 GB, and F32 0.010582016 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, BatiAI author metadata, and no MTP, mmproj, vision, or draft tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned Qwen base config, expanded linked-file size metadata, and a direct GGUF header/tensor-index range read of the selected IQ3_XXS artifact." }, "notes": "Use this profile for the BatiAI main IQ3_XXS GGUF text artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or MTP speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "bullpoint--qwen3-coder-next-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "bullpoint/Qwen3-Coder-Next-AWQ-4bit", "title": "Bullpoint Qwen3-Coder-Next AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the Bullpoint compressed-tensors AWQ 4-bit package of Qwen3-Coder-Next.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, pinned config comparison, compressed-tensors recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The API metadata records Qwen/Qwen3-Coder-Next as the quantized base model. Manual comparison found matching memory-relevant architecture fields between the Bullpoint config and the audited official BF16 base config: architecture, model type, layer count, full_attention_interval, hidden size, expert geometry, attention heads, KV heads, head dimensions, linear-attention state geometry, max positions, vocabulary size, and untied embeddings. The Bullpoint config adds compressed-tensors AWQ quantization metadata and omits torch_dtype while retaining dtype bfloat16." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 48.217683456, "main_resident_weight_gb": 47.5953536, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.08493696, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "resolved safetensors index total_size and direct shard-header stored bytes for the compressed-tensors AWQ package", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed BF16 language tensors plus expected distinct routed expert tensor groups", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. The compressed-tensors recipe excludes shared expert, shared expert gate, and router gate tensors from quantization, so they remain BF16 and are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "The AWQ package only quantizes routed expert Linear tensors. The recipe excludes embeddings, lm_head, norms, full attention, linear attention, router gates, shared expert gates, shared expert projections, and MTP paths. Header-derived stored bytes are therefore authoritative for the bound: routed expert tensors total 43.487723520 GB and divide exactly into 0.084936960 GB per expert index across 512 expert indexes; fixed ordinary text traffic totals 4.107630080 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing routed expert weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6051842087737511, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed I32 INT4 routed expert weights, BF16 scales, I64 shape metadata, and unquantized BF16 tensors. AWQ dequantization, activation traffic, router compute, expert compute, recurrent-state writes, and framework scheduling are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32 for targeted Linear tensors. Because attention, linear-attention, router, shared-expert, embedding, and head tensors are ignored by the recipe, the resident package is not an ideal flat 0.5 bytes per parameter. The profile keeps full-attention KV cache as BF16 because the config has no KV-cache quantization scheme." }, "evidence": [ { "label": "Bullpoint Qwen3-Coder-Next AWQ 4-bit API metadata", "url": "https://huggingface.co/api/models/bullpoint/Qwen3-Coder-Next-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "downloads", "base_model_proof", "license", "pipeline", "serving", "commit_sha" ], "notes": "At commit b035ac132045fd726eaa1ce53888df39cbb30291, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_next, qwen3-coder, qwen3-next, moe, awq, 4bit, int4, compressed-tensors, llm-compressor, vllm, sglang, endpoints_compatible, region:us, and base_model:Qwen/Qwen3-Coder-Next tags. Current downloads are 154238. The API config reports Qwen3NextForCausalLM/qwen3_next and compressed-tensors quantization metadata." }, { "label": "Bullpoint Qwen3-Coder-Next AWQ 4-bit config", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/raw/b035ac132045fd726eaa1ce53888df39cbb30291/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32." }, { "label": "Bullpoint Qwen3-Coder-Next AWQ 4-bit recipe", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/raw/b035ac132045fd726eaa1ce53888df39cbb30291/recipe.yaml", "source_type": "config", "supports": [ "quantization_scope", "serving", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe applies AWQ to Linear targets with 4-bit symmetric integer weights, group_size 32, group strategy, and MSE observer. It ignores model.embed_tokens, lm_head, norms, self_attn projections and norms, linear_attn projections/norm/conv/state scalars, mlp.gate, shared_expert_gate, shared_expert gate/up/down projections, and MTP paths. That leaves routed mlp.experts tensors as the quantized expert traffic." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no differences in the audited architecture fields between the Bullpoint AWQ config and the official BF16 base config. The Bullpoint repo adds compressed-tensors quantization_config and keeps dtype bfloat16 while preserving the base model geometry." }, { "label": "Bullpoint Qwen3-Coder-Next AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/resolve/b035ac132045fd726eaa1ce53888df39cbb30291/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The resolved safetensors index metadata reports total_size 48217683456 bytes. Direct range-read safetensors headers across all ten shards found 221847 tensors totaling the same 48.217683456 GB: I32 packed weights 38.654705664 GB, BF16 tensors/scales 9.561798144 GB, and I64 shape tensors 0.001179648 GB. Reconstructed logical parameters from packed I32 weights plus unquantized BF16 model tensors total 79.674391296B. The resident-only model.embed_tokens.weight tensor is 0.622329856 GB, leaving 47.595353600 GB main resident bytes. Routed expert tensors total 43.487723520 GB across 221184 tensors and divide exactly to 0.084936960 GB per expert index. Fixed ordinary text traffic, including linear attention, full attention, routers, shared expert, shared expert gates, norms, and lm_head, totals 4.107630080 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned config, compressed-tensors AWQ recipe, official BF16 base config comparison, resolved safetensors index, direct shard-header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the generated ideal 4-bit estimate, which missed that most always-on Qwen3-Next language tensors remain BF16, missed the shared expert, omitted fixed DeltaNet state, and understated both resident footprint and ordinary decode traffic." }, { "id": "calamitousfelicitousness--qwen2-5-32b-instruct-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic", "title": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the FP8-dynamic safetensors package of Qwen2.5 32B Instruct.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face API metadata, pinned served FP8 config, existing audited Qwen2.5 32B Instruct profile/config evidence, safetensors index, linked-object metadata, and direct safetensors header range reads", "config_compatible": true, "notes": "The model card metadata lists Qwen/Qwen2.5-32B as base_model, while the repo title and served config _name_or_path identify the artifact as Qwen2.5-32B-Instruct. The served FP8 config matches the audited Qwen/Qwen2.5-32B-Instruct geometry: Qwen2ForCausalLM, 64 layers, hidden size 5120, 40 attention heads, 8 KV heads, tie_word_embeddings false, use_sliding_window false, and 32768 max position embeddings." }, "architecture": { "canonical_architecture_id": "qwen2-5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 34.322132736, "swept_weight_gb": 32.764997376, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "base logical Qwen2.5 32B Instruct parameters with direct FP8/BF16/F32 safetensors stored-byte totals", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, lm_head.weight, and FP8 scale tensors", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 34.322132736 GB of tensor payload across seven shards. The package mixes F8_E4M3 matrix weights, BF16 input embeddings/lm_head/norms, and tiny F32 scale tensors. Logical parameter counts follow the audited Qwen2.5 32B Instruct BF16 profile so model identity remains the 32.763876352B architecture while weight traffic follows this quantized artifact's stored bytes." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served FP8 config records use_sliding_window false and sliding_window null. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served FP8 repo config. The model card describes longer-context support through YaRN, but the served config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0475601960909269, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-fp8-dynamic-qwen2.5-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored safetensors payload bytes: F8_E4M3 matrix weights, BF16 unquantized tensors, and F32 scale tensors. Dynamic activation quantization, dequantization, activation traffic, kernels, scheduler behavior, and cache writes are outside this memory-side bound.", "notes": "The served config records quant_method fp8, activation_scheme dynamic, and ignored_layers [lm_head]. The checkpoint also stores model.embed_tokens.weight as BF16; ordinary decode keeps that embedding resident but does not sweep it as a full matrix." }, "evidence": [ { "label": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic API metadata", "url": "https://huggingface.co/api/models/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 7ae45d106280752e6961841f0711344dedfcb2cd, the live API records a public Apache-2.0 text-generation safetensors repo with qwen2, chat, fp8, base_model Qwen/Qwen2.5-32B, region:us, and 153510 downloads. The API safetensors block reports BF16 1558254592, F8_E4M3 31205621760, and total 32763876352 storage-accounting elements." }, { "label": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic model card", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/raw/7ae45d106280752e6961841f0711344dedfcb2cd/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The card is the Qwen2.5-32B-Instruct card with Apache-2.0 licensing, Qwen2 architecture details, 64 layers, 40 Q heads, 8 KV heads, 32K served config context, and instructions for YaRN beyond 32768 tokens." }, { "label": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic config", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/raw/7ae45d106280752e6961841f0711344dedfcb2cd/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, torch_dtype bfloat16, 64 layers, hidden_size 5120, 40 attention heads, 8 KV heads, intermediate_size 27648, tie_word_embeddings false, use_sliding_window false, sliding_window null, max_window_layers 70, rms_norm_eps 1e-6, 32768 max position embeddings, and quantization_config with quant_method fp8, activation_scheme dynamic, and ignored_layers [lm_head]." }, { "label": "Qwen2.5 32B Instruct audited profile and config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_params_b", "embedding_layout", "kv_adapter" ], "notes": "The existing audited BF16 profile records the same served Instruct geometry and separates model.embed_tokens.weight as the resident-only input embedding while charging lm_head.weight as swept output-projection traffic." }, { "label": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic linked-object metadata", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/tree/7ae45d106280752e6961841f0711344dedfcb2cd", "source_type": "derived_calculation", "supports": [ "storage_overhead", "resident_weight_gb" ], "notes": "Expanded tree metadata records seven safetensors linked objects totaling 34.322264488 GB. The index records 34.322132736 GB of tensor payload, leaving 0.000131752 GB of safetensors header/container overhead outside tensor payloads." }, { "label": "CalamitousFelicitousness Qwen2.5 32B Instruct FP8 Dynamic safetensors headers", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/raw/7ae45d106280752e6961841f0711344dedfcb2cd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The index maps 1219 tensors across seven safetensors shards and records total_size 34322132736 bytes. Direct range-read headers match exactly: tensor payloads total 34.322132736 GB, split into F8_E4M3 31.205621760 GB, BF16 3.116509184 GB, and F32 0.000001792 GB. model.embed_tokens.weight is BF16 with shape [152064, 5120] and contributes 1.557135360 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.layers.* plus model.norm.weight plus lm_head.weight and scale tensors sum to 32.764997376 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned served FP8 config, existing audited Qwen2.5 32B Instruct profile/config evidence, safetensors index, expanded linked-object metadata, and direct safetensors shard header range reads." }, "notes": "Use this profile for the CalamitousFelicitousness FP8-dynamic safetensors artifact. Do not substitute AWQ/GPTQ or BF16 Qwen2.5 32B profiles; the tensor geometry is shared but the stored byte traffic and commit pin are distinct." }, { "id": "casperhansen--deepseek-r1-distill-llama-70b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "casperhansen/deepseek-r1-distill-llama-70b-awq", "title": "Casper Hansen DeepSeek R1 Distill Llama 70B AWQ", "summary": "Audited memory-side bounds profile for the Casper Hansen AWQ INT4 package of DeepSeek R1 Distill Llama 70B.", "model_family": "llama3.3-dense-awq", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "relation": "quantized", "source": "Served AWQ config _name_or_path, DeepSeek base snapshot config comparison, Hugging Face API metadata, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The served config _name_or_path identifies deepseek-ai/DeepSeek-R1-Distill-Llama-70B snapshot 07a264a567ba0863a4ab34fdb3c2b8a54e0bb494 as the source checkpoint. Manual comparison found matching architecture, context, attention, RoPE, vocabulary, and embedding-layout fields between that DeepSeek base snapshot and this AWQ config; the target adds AWQ quantization metadata and sets use_cache false as a generation default." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-llama-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 39.767785472, "swept_weight_gb": 37.666439168, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "logical DeepSeek R1 Distill Llama 70B parameters represented by safetensors qweight plus BF16/F16 tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors with F16 scales and unquantized BF16 embedding/head/norm tensors. Logical parameter counts use the API safetensors total and served tensor layout: I32 qweight tensors are unpacked 8x for logical weight count, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 80 layers, 8 KV heads, hidden size 8192, 64 attention heads, Llama 3 RoPE scaling, and no sliding-window setting. Although the package config and generation config set use_cache false by default, the audited Llama attention implementation supports standard K/V cache updates, so this profile charges full-context K and V streams for production cached text decode." }, "notes": "Dense LlamaForCausalLM AWQ profile using the served Casper Hansen repo config and tensor headers. The profile uses the public DeepSeek source snapshot rather than inferring from the model name." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5636526760540158, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-tgi-vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized BF16 tensors from safetensors headers. AWQ dequantization, Marlin/GEMM kernel differences, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by logical model parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Casper Hansen DeepSeek R1 Distill Llama 70B AWQ API metadata", "url": "https://huggingface.co/api/models/casperhansen/deepseek-r1-distill-llama-70b-awq", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "serving", "commit_sha", "logical_parameter_count" ], "notes": "At commit a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f, the live API records a public non-gated repo with safetensors, llama, MIT license, 4-bit, awq, and region:us tags. Current downloads are 423417. The API safetensors block reports I32 68451041280, BF16 2102665216, and total 70553706496 logical parameters." }, { "label": "Casper Hansen DeepSeek R1 Distill Llama 70B AWQ model card", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq/raw/a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f/README.md", "source_type": "model_card", "supports": [ "license" ], "notes": "The README contains MIT license frontmatter only. The executable artifact is therefore audited from its served config and tensor headers." }, { "label": "Casper Hansen DeepSeek R1 Distill Llama 70B AWQ served config", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq/raw/a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens", "base_model_proof" ], "notes": "The config records _name_or_path /root/.cache/huggingface/hub/models--deepseek-ai--DeepSeek-R1-Distill-Llama-70B/snapshots/07a264a567ba0863a4ab34fdb3c2b8a54e0bb494, LlamaForCausalLM, torch_dtype bfloat16, hidden_size 8192, intermediate_size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true. use_cache false is treated as a default generation setting, not a structural absence of K/V projections." }, { "label": "DeepSeek R1 Distill Llama 70B source snapshot config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/07a264a567ba0863a4ab34fdb3c2b8a54e0bb494/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in checked architecture, context, attention, RoPE, vocabulary, and embedding-layout fields between the served AWQ config and the source DeepSeek snapshot config. The source snapshot records torch_dtype bfloat16 and use_cache true." }, { "label": "DeepSeek R1 Distill Llama 70B source snapshot API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/revision/07a264a567ba0863a4ab34fdb3c2b8a54e0bb494", "source_type": "model_card", "supports": [ "base_model_proof", "logical_parameter_count" ], "notes": "The pinned source snapshot API records BF16 safetensors total 70553706496 parameters and region:us metadata." }, { "label": "Casper Hansen DeepSeek R1 Distill Llama 70B AWQ safetensors index and shard headers", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq/resolve/a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 39767785472 bytes across nine shards, matching direct range-read shard header spans. Headers contain 1843 tensors totaling 39.767785472 GB: I32 34.492907520 GB, BF16 4.205330432 GB, and F16 1.069547520 GB. Stored suffix totals are qweight 34.225520640 GB, qzeros 0.267386880 GB, scales 1.069547520 GB, and BF16 weight tensors 4.205330432 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 8192] and contribute 1.050673152B parameters / 2.101346304 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 37.666439168 GB. HEAD checks for all nine shards found linked sizes totaling 39767996432 bytes, leaving 210960 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Llama implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/llama/modeling_llama.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found LlamaModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 80 layers for this config. LlamaAttention projects key_states and value_states, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx), supporting expanded full-context BF16 K/V cache charges for ordinary cached decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, minimal model card frontmatter, served AWQ config, DeepSeek source snapshot API and config, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and the upstream Transformers Llama runtime implementation." }, "notes": "Use this profile for the Casper Hansen AWQ INT4 artifact only. It models production cached ordinary text decode, not the package's default use_cache false generation setting." }, { "id": "casperhansen--deepseek-r1-distill-qwen-32b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "casperhansen/deepseek-r1-distill-qwen-32b-awq", "title": "DeepSeek R1 Distill Qwen 32B AWQ", "summary": "Audited memory-side bounds profile for the Casper Hansen AWQ 4-bit DeepSeek R1 Distill Qwen 32B repo.", "model_family": "qwen2.5-dense-awq", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "relation": "quantized", "source": "Served AWQ config, audited BF16 DeepSeek profile comparison, safetensors tensor layout, and Hugging Face repo metadata", "config_compatible": true, "notes": "The repo metadata does not publish base_model card data and the README only declares MIT licensing. The served AWQ config, vocabulary, text geometry, untied embedding layout, and safetensors tensor names match the audited BF16 DeepSeek R1 Distill Qwen 32B profile for the fields used by this bounds profile, while adding AWQ GEMM 4-bit quantization. The AWQ config sets sliding_window null instead of the BF16 config's disabled sliding_window 131072; use_sliding_window is false in both, so the text KV adapter remains full-context." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-qwen2-5-32b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 19.328804864, "swept_weight_gb": 17.771669504, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus BF16 and F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, BF16 embedding/head/norm tensors, and F16 scales/biases. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros, scales, and biases are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The base BF16 config records sliding_window 131072, but sliding-window attention is disabled there too." }, "notes": "Dense Qwen2ForCausalLM AWQ profile using the served Casper Hansen repo config and tensor headers. The profile audits this served repo directly rather than deriving structure from the repo name." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5899425530831646, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, BF16/F16 scales, biases, and unquantized tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Casper Hansen DeepSeek R1 Distill Qwen 32B AWQ API metadata", "url": "https://huggingface.co/api/models/casperhansen/deepseek-r1-distill-qwen-32b-awq", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit e20a4933e66aa5eccc8270489f5aeab17f90b888, the live API reports a public non-gated repo with safetensors, qwen2, MIT license, 4-bit, awq, and region:us tags. Current downloads are 272214. The API safetensors block reports I32: 31205621760, BF16: 1557795840, F16: 458752, and total: 32763876352 logical parameters. The API does not publish pipeline_tag, library_name, or base_model card data for this repo." }, { "label": "Casper Hansen DeepSeek R1 Distill Qwen 32B AWQ README", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/raw/e20a4933e66aa5eccc8270489f5aeab17f90b888/README.md", "source_type": "model_card", "supports": [ "license", "metadata_limits" ], "notes": "The pinned README contains only frontmatter declaring MIT license, so architecture and base compatibility are audited from the served config and tensor layout rather than from narrative card claims." }, { "label": "Casper Hansen DeepSeek R1 Distill Qwen 32B AWQ served config", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/raw/e20a4933e66aa5eccc8270489f5aeab17f90b888/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with torch_dtype bfloat16, hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 131072, use_sliding_window false, sliding_window null, tie_word_embeddings false, vocab_size 152064, rope_theta 1000000, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "DeepSeek R1 Distill Qwen 32B BF16 profile and config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison against the already audited BF16 DeepSeek profile found matching checked architecture fields: Qwen2ForCausalLM, hidden size 5120, intermediate size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 131072, use_sliding_window false, tie_word_embeddings false, vocab size 152064, rope_theta 1000000, rms_norm_eps 1e-5, and untied input/lm_head tensors. The only checked config difference is sliding_window null here versus 131072 in the BF16 config, but it is disabled by use_sliding_window false in both configs." }, { "label": "Casper Hansen DeepSeek R1 Distill Qwen 32B AWQ safetensors index and shard headers", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/raw/e20a4933e66aa5eccc8270489f5aeab17f90b888/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 19328804864 bytes across four shards, matching direct range-read shard header spans. Headers contain 1667 tensors totaling 19.328804864 GB: 15.724707840 GB I32 tensors, 3.115591680 GB BF16 tensors, and 0.488505344 GB F16 tensors. Stored suffix totals are qweight 15.602810880 GB, qzeros 0.121896960 GB, scales 0.487587840 GB, bias 0.000917504 GB, and BF16 weight tensors 3.115591680 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 17.771669504 GB." }, { "label": "Casper Hansen DeepSeek R1 Distill Qwen 32B AWQ linked-object HEAD checks", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/tree/e20a4933e66aa5eccc8270489f5aeab17f90b888", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for all four safetensors shards found linked sizes 4956994936, 4993553856, 4960831008, and 4417614320 bytes. The linked file sizes include safetensors JSON header/container overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served AWQ config, minimal README, audited BF16 DeepSeek profile/config comparison, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and local scrape row." }, "notes": "Use this profile for the Casper Hansen AWQ artifact only. Do not substitute another Qwen2.5 32B AWQ profile even though the geometry and byte split match; each repo keeps its own evidence and commit pins." }, { "id": "casperhansen--llama-3-3-70b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "casperhansen/llama-3.3-70b-instruct-awq", "title": "Casper Hansen Llama 3.3 70B Instruct AWQ", "summary": "Audited memory-side bounds profile for the Casper Hansen AWQ INT4 Llama 3.3 70B Instruct repo.", "model_family": "llama3.3-dense-awq", "base_model_proof": { "base_model": "meta-llama/Llama-3.3-70B-Instruct", "relation": "quantized", "source": "Hugging Face model card text, served AWQ config _name_or_path, gated Llama 3.3 base-model API metadata, and safetensors header review", "config_compatible": false, "notes": "The repo name, README body, and served config _name_or_path identify this as an AWQ package of meta-llama/Llama-3.3-70B-Instruct. The README frontmatter and API cardData still list base_model meta-llama/Llama-3.1-70B, matching the upstream Meta lineage metadata rather than the served source checkpoint. Raw base config access remains gated in this audit environment, so direct config compatibility with the base cannot be independently verified. This profile audits the served AWQ artifact directly from its public config and tensor headers." }, "architecture": { "canonical_architecture_id": "llama-3-3-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 39.767785472, "swept_weight_gb": 37.666439168, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "logical Llama 3.3 70B parameters represented by safetensors qweight plus F16 tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors with F16 scales and unquantized F16 embedding/head/norm tensors. Logical parameter counts use the API safetensors total and the served tensor layout: I32 qweight tensors are unpacked 8x for logical weight count, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 80 layers, 8 KV heads, hidden size 8192, 64 attention heads, Llama 3 RoPE scaling, and no sliding-window setting, so this profile charges full-context K and V streams for ordinary cached text decode." }, "notes": "Dense LlamaForCausalLM AWQ profile using the served Casper Hansen repo config and tensor headers. The profile does not rely on direct access to the gated Meta base config." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5636526760540158, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-tgi-vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, Marlin/GEMM kernel differences, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by logical model parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Casper Hansen Llama 3.3 70B AWQ API metadata", "url": "https://huggingface.co/api/models/casperhansen/llama-3.3-70b-instruct-awq", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "serving", "commit_sha", "logical_parameter_count" ], "notes": "At commit 64d255621f40b42adaf6d1f32a47e1d4534c0f14, the live API records a public non-gated Transformers text-generation repo with safetensors, llama, text-generation-inference, endpoints_compatible, 4-bit, awq, deploy:azure, region:us, and Llama 3.3 license metadata. Current downloads are 318125. The API safetensors block reports I32 68451041280, F16 2102665216, total 70553706496 logical parameters. The API cardData base_model is meta-llama/Llama-3.1-70B, while the served config _name_or_path and README body identify the served source as Llama 3.3 70B Instruct." }, { "label": "Casper Hansen Llama 3.3 70B AWQ model card", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "quantization", "serving" ], "notes": "The README body says this is the AWQ version of the Llama 3.3 70B Instruct model and reproduces Meta's Llama 3.3 70B Instruct description, including 128k context, GQA, multilingual text in/out, and the Llama 3.3 license. Its YAML frontmatter still lists base_model meta-llama/Llama-3.1-70B, so this profile treats the served config and tensor headers as authoritative for the executable artifact." }, { "label": "Casper Hansen Llama 3.3 70B AWQ served config", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens", "base_model_proof" ], "notes": "The config records _name_or_path /root/.cache/huggingface/hub/models--meta-llama--Llama-3.3-70B-Instruct/snapshots/38ff4e01a70559264c95945aa04b900a11e68422, LlamaForCausalLM, torch_dtype float16, hidden_size 8192, intermediate_size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Meta Llama 3.3 70B Instruct base API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.3-70B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "logical_parameter_count", "license" ], "notes": "The current base API records gated manual access at commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b, text-generation pipeline, Llama 3.3 license, region:us, and BF16 safetensors total 70553706496 parameters. Raw base config and tensor index remain inaccessible in this audit environment, so the served AWQ config is the architecture source of truth." }, { "label": "Casper Hansen Llama 3.3 70B AWQ safetensors index and shard headers", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 39767785472 bytes across nine shards, matching direct range-read shard header spans. Headers contain 1843 tensors totaling 39.767785472 GB: I32 34.492907520 GB and F16 5.274877952 GB. Stored suffix totals are qweight 34.225520640 GB, qzeros 0.267386880 GB, scales 1.069547520 GB, and F16 weight tensors 4.205330432 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 8192] and contribute 1.050673152B parameters / 2.101346304 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 37.666439168 GB." }, { "label": "Casper Hansen Llama 3.3 70B AWQ linked-object HEAD checks", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/tree/64d255621f40b42adaf6d1f32a47e1d4534c0f14", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for all nine safetensors shards found linked sizes totaling 39767996256 bytes. The linked file sizes include safetensors JSON header/container overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served AWQ config, gated Llama 3.3 base API metadata, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and local scrape row." }, "notes": "Use this profile for the Casper Hansen AWQ INT4 artifact only. The executable artifact is audited directly and does not turn the gated Meta BF16 base profile into an audited profile." }, { "id": "chunity--gemma-4-e4b-it-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Chunity/gemma-4-E4B-it-AWQ-4bit", "title": "Chunity Gemma 4 E4B IT AutoRound AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for Chunity's AutoRound AWQ 4-bit package of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, model card, quantization config, and direct safetensors header grouping", "config_compatible": false, "notes": "The repo metadata records google/gemma-4-E4B-it as its quantized base model. Manual comparison found matching checked text, vision, audio, context, tied-embedding, and attention geometry fields between the Chunity config and the pinned Google base config. The served Chunity config adds AutoRound AWQ metadata and changes top-level dtype fields to float16 while retaining bfloat16 text_config dtype, so this profile uses the served repo config directly for serving formats." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.805237482, "swept_params_b": 1.50857681, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 10.522597844, "swept_weight_gb": 3.9292765, "auxiliary_resident_weight_gb": 6.593321344, "resident_parameter_scope": "safetensors_header_stored_awq_i32_bf16_f16_tensor_elements", "swept_parameter_scope": "ordinary text decode includes model.language_model tensors excluding the per-layer embedding table but including the tied standard embedding/output projection", "auxiliary_scope": "model.language_model.embed_tokens_per_layer.weight, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision are resident for multimodal PLE packaging but not swept as full matrices for each ordinary generated text token", "notes": "Header-derived bytes are used because this AutoRound AWQ artifact mixes packed I32 int4 language tensors, F16 scales, BF16 embeddings, BF16 higher-precision attention and multimodal tensors, and BF16 side tensors. The config records tie_word_embeddings true and the headers have no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The large per-layer embedding table remains resident-only, matching the audited Google Gemma 4 E4B PLE convention." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, with separate K and V streams." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 35 layers use 512-token local sliding-window attention with separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The quantization config has no kv_cache_scheme, so this profile charges two-byte KV cache streams. The top-level runtime dtype is float16; the text_config dtype is bfloat16. Both are two-byte cache formats, and audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 2.189818481067113, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoround-awq-int4-fp16-gemma4-e4b-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored safetensors bytes: packed I32 AWQ tensors, F16 scale tensors, BF16 preserved weights, BF16 embeddings, BF16 multimodal tensors, and tiny BF16 side tensors from the shard headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records AutoRound AWQ 4-bit integer weights, group_size 128, symmetric quantization, zero_point false, and quantized block model.language_model.layers. modules_to_not_convert preserves vision_tower, audio_tower, embed_vision, embed_audio, lm_head, and selected attention modules in higher precision. The README validated the Transformers AWQ loader with dtype auto." }, "evidence": [ { "label": "Chunity Gemma 4 E4B AWQ API metadata", "url": "https://huggingface.co/api/models/Chunity/gemma-4-E4B-it-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700, the API records a public non-gated Transformers image-text-to-text repo with base_model google/gemma-4-E4B-it, base_model:quantized metadata, awq, autoround, 4-bit, multimodal, endpoints_compatible, and region:us tags. Current downloads are 223,205. The API safetensors block reports I32 3,620,208,640 and BF16 4,320,892,234 entries, while the safetensors index records 4,805,237,482 stored tensor elements; this profile uses direct shard headers for exact stored bytes and scope splits." }, { "label": "Chunity Gemma 4 E4B AWQ model card", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "quantized_module_scope" ], "notes": "The README frontmatter records base_model google/gemma-4-E4B-it and quantized_by Chunity. The visible card describes an AutoRound AWQ 4-bit quantization with group size 128, 500 iterations, quantized block model.language_model.layers, and higher-precision preservation for vision_tower, audio_tower, embed_vision, embed_audio, and lm_head. It says the checkpoint was smoke-tested with the Transformers AWQ loader." }, { "label": "Chunity Gemma 4 E4B AWQ config", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit/raw/cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, AutoRound AWQ 4-bit integer quantization with group_size 128, symmetric quantization, zero_point false, quantized block model.language_model.layers, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Chunity Gemma 4 E4B AWQ quantization config", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit/raw/cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700/quantization_config.json", "source_type": "config", "supports": [ "serving", "quantization", "quantized_module_scope" ], "notes": "The standalone quantization_config records AutoRound version 0.12.2, provider auto-round, quant_method awq, bits 4, group_size 128, sym true, zero_point false, seqlen 256, nsamples 32, iters 500, and modules_to_not_convert entries that preserve multimodal towers, embeddings, lm_head, and selected attention modules." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no profile-relevant geometry differences between the Chunity config and the pinned Google base config after excluding quantization metadata, dtype fields, and repository bookkeeping. The checked text layer pattern, hidden size, KV heads, head dimensions, context, tied embeddings, vision config, and audio config match." }, { "label": "Chunity Gemma 4 E4B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit/raw/cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The index maps 2,656 tensors across seven shards and records metadata.total_size 10.522597844 GB. Range-reading every shard header found exactly 10.522597844 GB of direct tensor payload: BF16 8.641786324 GB, I32 1.824245760 GB, and F16 0.056565760 GB. Linked-object HEAD checks resolved the seven shards to 10.522940460 GB, leaving 342,616 bytes of safetensors header/container overhead outside tensor payloads. model.language_model.embed_tokens.weight is BF16 [262144, 2560] and contributes 1.342177280 GB; because there is no separate lm_head.weight, this tied output projection is swept for ordinary decode. model.language_model.embed_tokens_per_layer.weight is BF16 [262144, 10752] and contributes 5.637144576 GB resident-only for ordinary decode. Audio/embed_audio tensors total 0.617514496 GB, and vision/embed_vision tensors total 0.338662272 GB. Ordinary text swept traffic, defined as language_model tensors excluding only embed_tokens_per_layer, totals 3.929276500 GB. Auxiliary resident tensors, defined as per-layer embedding plus audio plus vision, total 6.593321344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served AutoRound AWQ config, standalone quantization config, pinned Google base config comparison, linked-object HEAD checks, and direct seven-shard safetensors header byte grouping." }, "notes": "This profile supersedes the generated metadata estimate, which treated the artifact as an ideal 0.5-byte dense model and missed BF16 preserved weights, F16 AWQ scale tensors, multimodal towers, per-layer embedding residency, tied output projection traffic, and hybrid Gemma 4 KV traffic." }, { "id": "coherelabs--aya-vision-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "CohereLabs/aya-vision-8b", "title": "CohereLabs Aya Vision 8B F16", "summary": "Unsupported profile stub for the gated F16 Aya Vision 8B multimodal repo.", "model_family": "aya-vision-multimodal", "architecture": { "canonical_architecture_id": "aya-vision-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.631842032, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8631842032 F16 safetensors parameters for this repo. Text KV geometry, image adapter geometry, and resident-only multimodal tensors are not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw README, and raw safetensors index returned access denied in this audit environment.", "notes": "Do not infer Aya Vision text layers, KV heads, exact head dimension, image-token behavior, multimodal projector residency, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and multimodal resident/swept scope are unavailable.", "notes": "F16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "CohereLabs Aya Vision 8B API metadata", "url": "https://huggingface.co/api/models/CohereLabs/aya-vision-8b", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit d7601ac8fd3baf6f59c5fc9514c618ef0f24268f, the live API reports gated: auto, image-text-to-text pipeline, cc-by-nc-4.0 license, 161292 downloads, region:us, and F16 safetensors count 8631842032." }, { "label": "CohereLabs Aya Vision 8B gated config and index access check", "url": "https://huggingface.co/CohereLabs/aya-vision-8b/raw/d7601ac8fd3baf6f59c5fc9514c618ef0f24268f/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config.json, README.md, and model.safetensors.index.json requests returned HTTP 401 access-restricted responses in this audit environment." } ], "unsupported_reason": "Gated config and tensor index are not accessible in this audit environment, so text KV geometry, multimodal residency, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from current HF API metadata and direct raw config/README/safetensors-index access checks." }, "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "cortecs--llama-3-3-70b-instruct-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic", "title": "cortecs Llama 3.3 70B Instruct FP8 Dynamic", "summary": "Audited memory-side bounds profile for the cortecs compressed-tensors FP8 dynamic package of Llama 3.3 70B Instruct.", "model_family": "llama3.3-dense-fp8", "base_model_proof": { "base_model": "meta-llama/Llama-3.3-70B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served compressed-tensors config, gated base API metadata, audited Llama 3.3 derivative profiles, and safetensors header review", "config_compatible": true, "notes": "The repo card and API identify this as a quantized derivative of meta-llama/Llama-3.3-70B-Instruct. Raw base config access remains gated in this audit environment, but the public cortecs config directly records the Llama 3.3 70B architecture and matches the memory-relevant geometry used by already audited Llama 3.3 70B derivative profiles." }, "architecture": { "canonical_architecture_id": "llama-3-3-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.560423936, "swept_params_b": 69.509750784, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 72.669806592, "swept_weight_gb": 70.568460288, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "safetensors_header_stored_fp8_f16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, lm_head.weight, and compressed-tensors scale tensors", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The config records tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight F16 tensors. The API safetensors total counts F16 and F8_E4M3 storage-accounting tensor elements, including per-channel F16 weight_scale tensors; direct safetensors data_offsets drive byte traffic." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 80 layers, 8 KV heads, 128 head dimension, 131072 max position embeddings, and no sliding-window setting, so this profile charges full-context K and V streams for ordinary cached text decode." }, "notes": "Dense LlamaForCausalLM FP8 dynamic profile using the served cortecs repo config and exact stored safetensors bytes. The profile does not rely on direct access to the gated Meta base config." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus F16 embeddings, lm_head, norms, and per-channel weight_scale tensors from safetensors headers. Dynamic activation quantization, dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and compressed-tensors float-quantized FP8 weights with dynamic token activation quantization, static channel weight quantization, and lm_head ignored. kv_cache_scheme is null, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "cortecs Llama 3.3 70B FP8 Dynamic API metadata", "url": "https://huggingface.co/api/models/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "serving", "commit_sha", "logical_parameter_count" ], "notes": "At commit 3722358cc2b990b22304489b2f87ef3bb876d6f6, the live API records a public non-gated repo with safetensors, llama, compressed-tensors, region:us, license:other, base_model meta-llama/Llama-3.3-70B-Instruct, and 145566 downloads. The API safetensors block reports F16 2109382656, F8_E4M3 68451041280, and total 70560423936 storage-accounting tensor elements." }, { "label": "cortecs Llama 3.3 70B FP8 Dynamic model card", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic/raw/3722358cc2b990b22304489b2f87ef3bb876d6f6/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "quantization", "serving" ], "notes": "The pinned card identifies this as a quantization of Llama-3.3-70B-Instruct, reports 99.67% accuracy recovery in its evaluation summary, provides vLLM serving guidance, and links the Llama license." }, { "label": "cortecs Llama 3.3 70B FP8 Dynamic served config", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic/raw/3722358cc2b990b22304489b2f87ef3bb876d6f6/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens", "base_model_proof" ], "notes": "The config records LlamaForCausalLM, torch_dtype float16, hidden_size 8192, intermediate_size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and compressed-tensors FP8 dynamic activation quantization with lm_head ignored and kv_cache_scheme null." }, { "label": "Meta Llama 3.3 70B Instruct base API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.3-70B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "logical_parameter_count", "license" ], "notes": "The current base API records gated manual access at commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b, text-generation pipeline, Llama 3.3 license, region:us, and BF16 safetensors total 70553706496 parameters. Raw base config and tensor index remain inaccessible in this audit environment, so the served cortecs config is the architecture source of truth." }, { "label": "cortecs Llama 3.3 70B FP8 Dynamic safetensors index and shard headers", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic/raw/3722358cc2b990b22304489b2f87ef3bb876d6f6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 72669806592 bytes across 15 shards, matching direct range-read shard header spans. Headers contain 1283 tensors totaling 72.669806592 GB: F8_E4M3 68.451041280 GB and F16 4.218765312 GB. Stored suffix totals are weight 72.656371712 GB and weight_scale 0.013434880 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 8192] and contribute 1.050673152B tensor elements / 2.101346304 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 70.568460288 GB. Linked-object HEAD checks resolved all shards to 72.669953960 GB, leaving 147368 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Audited Llama 3.3 70B derivative geometry comparison", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison against already audited public Llama 3.3 derivative configs found matching memory-relevant geometry: LlamaForCausalLM, hidden size 8192, intermediate size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, and Llama 3 RoPE scaling." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, served FP8 dynamic config, gated Llama 3.3 base API metadata, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and audited public Llama 3.3 derivative config comparison." }, "notes": "Use this profile for the cortecs FP8 dynamic artifact only. The executable artifact is audited directly and does not turn the gated Meta BF16 base profile into an audited profile." }, { "id": "cyankiwi--gemma-4-12b-it-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/gemma-4-12B-it-AWQ-INT4", "title": "cyankiwi Gemma 4 12B IT AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ INT4 package of Gemma 4 12B Unified IT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": false, "notes": "The AWQ repo records google/gemma-4-12B-it as its base model. Manual comparison found matching Gemma4UnifiedForConditionalGeneration shape, text layer geometry, local/global attention pattern, tied embeddings, dense setting, and lightweight unified multimodal geometry. The served AWQ config changes text_config.max_position_embeddings from 262144 to 131072 and text_config.dtype from the BF16 base to float32, so this profile uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 12.6409776, "swept_params_b": 12.588597696, "auxiliary_resident_params_b": 0.052379904, "resident_weight_gb": 11.221914944, "swept_weight_gb": 11.012395328, "auxiliary_resident_weight_gb": 0.209519616, "resident_parameter_scope": "hf_api_logical_compressed_tensors_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "model.language_model safetensors headers, including packed AWQ tensors, F32 scales/zero points, I64 shape side tensors, F32 norms, and the tied embedding/output projection", "auxiliary_scope": "model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for multimodal inputs but not swept for each generated text token", "notes": "HF API logical compressed-tensors parameters are used for parameter counts, with I32 packed tensors counted as unpacked 4-bit logical weights. Exact range-read safetensors header bytes drive memory footprint and per-token traffic. The served config records tie_word_embeddings true and the header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.8877410671149358, "kv_store_format": "fp32", "kv_store_bytes_per_scalar": 4, "kv_read_format": "fp32", "kv_read_bytes_per_scalar": 4, "runtime_format": "transformers-compressed-tensors-awq-fp32-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored compressed-tensors bytes: packed I32 int4 payloads, F32 scale and zero-point tensors, ignored F32 modules, and I64 shape side tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, asymmetric group quantization, no KV cache quantization scheme, top-level dtype bfloat16, and text_config dtype float32. This profile charges FP32 KV/state scalars from the served text_config dtype." }, "evidence": [ { "label": "cyankiwi Gemma 4 12B AWQ INT4 model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-12B-it-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The live HF API response records repo SHA aa32360daab742921c34e01a503c024fc8c6a510, Apache-2.0 licensing, any-to-any pipeline, region:us tag, base_model google/gemma-4-12B-it, compressed-tensors metadata, 364002 downloads, and safetensors logical parameters I64: 656, F32: 1400406064, I32: 11240570880, total: 12640977600. The model card describes cyankiwi AWQ version 26.05.01 with STEM and Agentic calibration and an 11.22 GB model size." }, { "label": "cyankiwi Gemma 4 12B AWQ INT4 config", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4/raw/aa32360daab742921c34e01a503c024fc8c6a510/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, compressed-tensors pack-quantized int4 weights, group_size 32, asymmetric group strategy, no KV cache scheme, tie_word_embeddings true, top-level dtype bfloat16, text_config dtype float32, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 131072 max position embeddings, 16 attention heads, 8 sliding KV heads, 1 global KV head, 256 sliding head dimension, 512 global head dimension, and lightweight audio/vision projection configs." }, { "label": "Google Gemma 4 12B IT base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found matching checked architecture fields except text dtype and context cap, plus non-decode audio/vision config default-field differences. The base config records BF16 text dtype and 262144 max positions; this AWQ artifact records float32 text dtype and 131072 max positions." }, { "label": "cyankiwi Gemma 4 12B AWQ INT4 README", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "context", "unified_multimodal" ], "notes": "The README frontmatter records base_model google/gemma-4-12B-it and Apache-2.0 licensing. The visible model card describes the 12B Unified encoder-free multimodal model, local/global hybrid attention, 1024-token sliding window, and the Gemma 4 family context-window claims; the served config remains authoritative for this artifact's 131072-token cap." }, { "label": "cyankiwi Gemma 4 12B AWQ INT4 safetensors header", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4/resolve/aa32360daab742921c34e01a503c024fc8c6a510/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The linked object size is 11222132768 bytes, with a 217816-byte header and 11221914944 tensor bytes across 1661 tensors. Stored payloads are I32 5.620285440 GB, F32 5.601624256 GB, and I64 0.000005248 GB. Logical compressed-tensors parameters are 12.640977600B after counting I32 packed int4 payloads as unpacked logical weights. model.language_model tensors total 12.588597696B logical parameters / 11.012395328 GB and are swept for ordinary text decode, including the tied embedding/output projection. Resident-only multimodal tensors outside model.language_model total 0.052379904B parameters / 0.209519616 GB. The header has model.language_model.embed_tokens.weight and no separate lm_head.weight." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, HF CLI repo info/file listing, the model card, pinned served compressed-tensors config, current base config comparison, linked object metadata, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, including the generated ideal 6.3205 GB resident estimate, flat full-context KV estimate, and stale download count." }, { "id": "cyankiwi--gemma-4-26b-a4b-it-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit", "title": "cyankiwi Gemma 4 26B A4B IT AWQ 4-bit", "summary": "Audited memory-side bounds profile for the cyankiwi AWQ 4-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records google/gemma-4-26B-A4B-it as its base model. Manual comparison found matching text and vision geometry: 30 language layers, hybrid local/global attention, 1024-token sliding window, attention_k_eq_v true, 128 experts, 8 routed experts per token, tied embeddings, resident vision tower, and 262144 max context." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.186571212, "main_resident_weight_gb": 16.04098238, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 3.19443134, "routed_expert_weight_gb": 0.10036368, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_f16_i64", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The base Gemma card describes 1 shared expert and the config records top_k_experts 8. The AWQ quantization config ignores model.language_model.layers.*.mlp.* and router.proj, so shared/always-on MLP and router tensors remain F16 and are charged in fixed_weight_gb.", "notes": "Header-derived bytes are used because the AWQ package stores packed I32 tensors plus F16 scale tensors, I64 shape side tensors, and unquantized F16 modules. Routed expert tensors are byte-uniform across all 128 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile models ordinary text decode for the AWQ package. Vision prefill and multimodal processor costs are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, F16 scales, I64 shape side tensors, and unquantized F16 modules from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and compressed-tensors pack-quantized int4 weights with group size 32. KV cache is charged at two bytes per scalar, matching the base Gemma profiles." }, "evidence": [ { "label": "cyankiwi Gemma 4 26B A4B IT AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "The API records repo SHA 4033b16200f4152e55e100ea12dc388c537df622, Apache-2.0 licensing, image-text-to-text pipeline, base_model:google/gemma-4-26B-A4B-it, and safetensors dtypes I64, I32, and F16. The card identifies the package as cyankiwi AWQ version 26.05.01 with STEM and Agentic calibration." }, { "label": "cyankiwi Gemma 4 26B A4B IT AWQ config", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit/raw/4033b16200f4152e55e100ea12dc388c537df622/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors pack-quantized int4 weights, group_size 32, symmetric group strategy, bfloat16 dtype, tie_word_embeddings true, attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, intermediate_size 2112, moe_intermediate_size 704, and 262144 max position embeddings. The quantization ignore list has 312 entries: 90 language MLP modules, 30 router projections, 190 vision modules, model.embed_vision, and lm_head." }, { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The base card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo and this AWQ artifact; the AWQ artifact adds quantization_config while preserving the base architecture." }, { "label": "cyankiwi Gemma 4 26B A4B IT AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit/raw/4033b16200f4152e55e100ea12dc388c537df622/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 17186571212 bytes across four shards. Range-read safetensors headers found 35743 tensors totaling 17.186571212 GB: 11.97408256 GB I32 packed tensors, 5.212302492 GB F16 tensors, and 0.00018616 GB I64 shape tensors. Resident language tensors total 16.04098238 GB; resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic, defined as non-expert language tensors including tied model.language_model.embed_tokens.weight, totals 3.19443134 GB. Routed expert tensors total 12.84655104 GB and divide exactly into 128 uniform expert groups of 0.10036368 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from cyankiwi and Google model cards, served config, base config comparison, HF API metadata, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this AWQ repo, including the catalog row's incorrect routed-experts-per-token value." }, { "id": "cyankiwi--gemma-4-26b-a4b-it-awq-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit", "title": "cyankiwi Gemma 4 26B A4B IT AWQ 8-bit", "summary": "Audited memory-side bounds profile for the cyankiwi AWQ 8-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records google/gemma-4-26B-A4B-it as its base model. Manual comparison against the audited cyankiwi 4-bit sibling and base geometry found matching text and vision architecture: 30 language layers, hybrid local/global attention, 1024-token sliding window, attention_k_eq_v true, 128 experts, 8 routed experts per token, tied embeddings, resident vision tower, and 262144 max context. The audited config-field difference from the 4-bit sibling is compressed-tensors AWQ weight bit width: 8 instead of 4." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 29.160467612, "main_resident_weight_gb": 28.01487878, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 3.74951942, "routed_expert_weight_gb": 0.18957312, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_int8_i32_bf16", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The base Gemma card describes 1 shared expert and the config records top_k_experts 8. The AWQ quantization config ignores model.language_model.layers.*.mlp.* and router.proj, so shared/always-on MLP and router tensors remain BF16 and are charged in fixed_weight_gb.", "notes": "Header-derived bytes are used because the AWQ package stores packed I32 tensors plus BF16 scale tensors and unquantized BF16 modules. Routed expert tensors are packed as grouped [128, ...] tensors; the profile divides total routed expert bytes by 128, matching the audited 4-bit sibling." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile models ordinary text decode for the AWQ package. Vision prefill and multimodal processor costs are outside Bounds Engine v1." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, BF16 scales, and unquantized BF16 modules from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and compressed-tensors pack-quantized int8 weights with group size 32. KV cache is charged at two bytes per scalar, matching the base Gemma profiles." }, "evidence": [ { "label": "cyankiwi Gemma 4 26B A4B IT AWQ 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "The API records repo SHA 7d086af42fd9ee78a877a1206b28dedd4f99e5e6, Apache-2.0 licensing, image-text-to-text pipeline, region:us, base_model:google/gemma-4-26B-A4B-it, and safetensors dtypes I32 and BF16. Current downloads were 169666 at audit time. API safetensors metadata reports I32 23948165120 parameters, BF16 2606151246 parameters, and 26554316366 total logical parameters." }, { "label": "cyankiwi Gemma 4 26B A4B IT AWQ 8-bit config", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit/raw/7d086af42fd9ee78a877a1206b28dedd4f99e5e6/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors pack-quantized int8 weights, group_size 32, symmetric group strategy, bfloat16 dtype, tie_word_embeddings true, attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, intermediate_size 2112, moe_intermediate_size 704, and 262144 max position embeddings. The quantization ignore list has the same 312 entries as the audited 4-bit sibling: 90 language MLP modules, 30 router projections, 190 vision modules, model.embed_vision, and lm_head." }, { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The base card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT base config and audited 4-bit sibling comparison", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo, the audited cyankiwi AWQ 4-bit artifact, and this AWQ 8-bit artifact. The AWQ artifacts add quantization_config while preserving the base architecture." }, { "label": "cyankiwi Gemma 4 26B A4B IT AWQ 8-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit/raw/7d086af42fd9ee78a877a1206b28dedd4f99e5e6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 29160467612 bytes across six shards. Range-read safetensors headers found 1188 tensors totaling 29.160467612 GB: 23.948165120 GB I32 packed tensors and 5.212302492 GB BF16 tensors. Resident language tensors total 28.014878780 GB; resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic, defined as non-expert language tensors including tied model.language_model.embed_tokens.weight, totals 3.749519420 GB. Routed expert tensors total 24.265359360 GB and divide into 128 uniform expert groups of 0.189573120 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from cyankiwi and Google model cards, served config, audited 4-bit sibling comparison, HF API metadata, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this AWQ repo, including the catalog row's incorrect 4-bit label and routed-experts-per-token value." }, { "id": "cyankiwi--gemma-4-31b-it-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/gemma-4-31B-it-AWQ-4bit", "title": "cyankiwi Gemma 4 31B IT AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi AWQ 4-bit package of Gemma 4 31B IT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records google/gemma-4-31B-it as its base model. Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, dense setting, and vision geometry between the cyankiwi AWQ artifact and the Google BF16 base repo. The AWQ artifact changes dtype labels to float16 and adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.103510176, "swept_params_b": 32.52776664, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 20.904118392, "swept_weight_gb": 19.75263132, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "hf_api_logical_compressed_tensors_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "model.language_model safetensors headers, including packed AWQ tensors, F16 scales, I64 shape side tensors, unquantized F16 tensors, norm, and tied embedding/output projection", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary generated text tokens", "notes": "HF API logical compressed-tensors parameters are used for parameter counts, while exact range-read safetensors header bytes drive memory footprint and per-token traffic. The served config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6314773956254179, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored AWQ packed weights, F16 scales, I64 shape side tensors, and unquantized F16 modules from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32 and no KV cache quantization scheme, so KV is charged at two bytes per scalar like the base BF16 Gemma profiles." }, "evidence": [ { "label": "cyankiwi Gemma 4 31B IT AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-31B-it-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f, the API records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-31B-it, endpoints_compatible and region:us tags, 1,035,024 downloads, and safetensors logical parameters I64: 820, I32: 30201937920, F16: 2901571436, total: 33103510176." }, { "label": "cyankiwi Gemma 4 31B IT AWQ served config", "url": "https://huggingface.co/cyankiwi/gemma-4-31B-it-AWQ-4bit/raw/853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors pack-quantized int4 weights, group_size 32, asymmetric group strategy, no KV cache scheme, tie_word_embeddings true, float16 text dtype, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 262144 max position embeddings, and a resident vision tower. The quantization ignore list has 192 entries: 190 vision modules, model.embed_vision, and lm_head." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this cyankiwi AWQ artifact; the cyankiwi artifact adds quantization_config while preserving the base architecture." }, { "label": "Google Gemma 4 31B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it", "source_type": "derived_calculation", "supports": [ "base_model_proof", "base_model_safetensors" ], "notes": "At commit 3548789868c5356dbf307c98e6f609007b82b3eb, the base API records image-text-to-text packaging, Apache-2.0 licensing, and safetensors BF16 parameters 31273088876 with total_size 32682372656." }, { "label": "cyankiwi Gemma 4 31B IT AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/gemma-4-31B-it-AWQ-4bit/raw/853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records total_size 20904118392 bytes across four shards. Range-read safetensors headers found 2418 tensors totaling 20.904118392 GB: 15.10096896 GB I32 packed tensors, 5.803142872 GB F16 tensors, and 0.00000656 GB I64 shape tensors. Language tensors under model.language_model total 6.10107096 stored tensor params / 19.75263132 GB and are swept for ordinary text decode, including the tied embedding/output projection. Resident-only vision tensors under model.vision_tower plus model.embed_vision total 0.575743536 stored params / 1.151487072 GB. The index has no separate lm_head.weight." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from cyankiwi and Google model metadata, pinned served configs, base config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, including the generated 16.5517 GB resident estimate and flat full-context KV estimate." }, { "id": "cyankiwi--gemma-4-e4b-it-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/gemma-4-E4B-it-AWQ-INT4", "title": "cyankiwi Gemma 4 E4B IT AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ INT4 package of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct safetensors header grouping", "config_compatible": false, "notes": "The repo metadata records google/gemma-4-E4B-it as its quantized base model. Manual comparison found matching checked text, vision, audio, context, tied-embedding, and attention geometry fields between the AWQ config and the pinned Google base config. The served AWQ config changes top-level, text, vision, and audio dtype fields from bfloat16 to float16, so this profile uses the served repo config directly for serving formats." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.185960272, "swept_params_b": 4.8892996, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 10.311668228, "swept_weight_gb": 3.718346884, "auxiliary_resident_weight_gb": 6.593321344, "resident_parameter_scope": "hf_api_logical_compressed_tensors_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "ordinary text decode includes model.language_model tensors excluding the per-layer embedding table but including the tied standard embedding/output projection", "auxiliary_scope": "model.language_model.embed_tokens_per_layer.weight, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision are resident for multimodal PLE packaging but not swept as full matrices for each ordinary generated text token", "notes": "HF API logical compressed-tensors parameters are used for parameter counts, with I32 packed int4 tensors counted as unpacked logical weights. Exact range-read safetensors header bytes drive resident memory and per-token swept traffic. The served config records tie_word_embeddings true and the header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The large per-layer embedding table remains resident-only, matching the audited Google Gemma 4 E4B PLE convention." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, with separate K and V streams." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 35 layers use 512-token local sliding-window attention with separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The compressed-tensors config has kv_cache_scheme null, so this profile charges FP16 KV cache bytes from the served dtype. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.2596772871315982, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-fp16-gemma4-e4b-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored compressed-tensors bytes: packed I32 int4 payloads, F16 scale, zero-point, norm, embedding, multimodal, and side tensors, plus I64 shape side tensors from the safetensors header. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, asymmetric group quantization, quantized Linear targets, and kv_cache_scheme null. The served top-level, text, vision, and audio dtype fields are float16." }, "evidence": [ { "label": "cyankiwi Gemma 4 E4B AWQ INT4 API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-E4B-it-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At repo SHA fc24b63a5f3313720e2d0adcdb915d7f3dc91392, the API records a public non-gated Apache-2.0 any-to-any repo with base_model google/gemma-4-E4B-it, base_model:quantized metadata, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 247056. The API safetensors block reports I64: 518, I32: 4040171520 logical packed int4 parameters, F16: 4145788234, and total: 8185960272 logical parameters." }, { "label": "cyankiwi Gemma 4 E4B AWQ INT4 model card", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "context", "multimodal_ple" ], "notes": "The README frontmatter records base_model google/gemma-4-E4B-it and Apache-2.0 licensing. The visible model card describes cyankiwi AWQ version 26.05.01, STEM and Agentic calibration, a 10.31 GB model size, and repeats the Gemma 4 E4B architecture summary: 4.5B effective parameters, 8B with embeddings, 42 layers, 512-token sliding window, 128K context, and text/image/audio support." }, { "label": "cyankiwi Gemma 4 E4B AWQ INT4 config", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4/raw/fc24b63a5f3313720e2d0adcdb915d7f3dc91392/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors AWQ pack quantization with 4-bit int weights, group_size 32, asymmetric group strategy, quantized Linear targets, kv_cache_scheme null, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no profile-relevant geometry differences between the cyankiwi AWQ config and the pinned Google base config after excluding quantization metadata and repository bookkeeping. The checked dtype fields differ: the Google base is bfloat16, while this served AWQ config is float16." }, { "label": "cyankiwi Gemma 4 E4B AWQ INT4 safetensors header", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4/resolve/fc24b63a5f3313720e2d0adcdb915d7f3dc91392/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The linked object is 10.312042292 GB with a 374056-byte safetensors header. Range-reading the single header found 2853 tensors with payloads totaling 10.311668228 GB: F16 8.291578324 GB, I32 2.020085760 GB, and I64 0.000004144 GB. model.language_model.embed_tokens.weight is F16 [262144, 2560] and contributes 1.342177280 GB; because there is no separate lm_head.weight, this tied output projection is swept for ordinary decode. model.language_model.embed_tokens_per_layer.weight is F16 [262144, 10752] and contributes 5.637144576 GB resident-only for ordinary decode. Audio/embed_audio tensors total 0.617514496 GB, and vision/embed_vision tensors total 0.338662272 GB. Ordinary text swept traffic, defined as language_model tensors excluding only embed_tokens_per_layer, totals 3.718346884 GB. Auxiliary resident tensors, defined as per-layer embedding plus audio plus vision, total 6.593321344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served compressed-tensors config, pinned Google base config comparison, linked-object HEAD check, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the generated metadata estimate, which treated the artifact as an ideal 0.5-byte dense model and missed F16 embeddings, F16 AWQ side tensors, multimodal towers, per-layer embedding residency, tied output projection traffic, and hybrid Gemma 4 KV traffic." }, { "id": "cyankiwi--glm-4-5-air-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/GLM-4.5-Air-AWQ-4bit", "title": "cyankiwi GLM 4.5 Air AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi AWQ 4-bit GLM-4.5-Air repo.", "model_family": "glm4-moe", "base_model_proof": { "base_model": "zai-org/GLM-4.5-Air", "relation": "quantized", "source": "Hugging Face model metadata, model card, served config, and base config comparison", "config_compatible": true, "notes": "The repo metadata records base_model zai-org/GLM-4.5-Air and base_model:quantized:zai-org/GLM-4.5-Air. Manual comparison found the served AWQ config preserves the base GLM-4.5-Air tensor geometry, MoE routing fields, attention geometry, context length, and untied embedding setting while adding compressed-tensors AWQ metadata." }, "architecture": { "canonical_architecture_id": "glm-4-5-air", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 63.394045248, "main_resident_weight_gb": 62.152531264, "auxiliary_resident_weight_gb": 1.241513984, "fixed_weight_gb": 6.095381824, "routed_expert_weight_gb": 0.43794648, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_bf16_f32_i64", "traffic_scope": "ordinary text decode through model.layers and lm_head, excluding resident-only input embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1, and the quantization ignore list leaves shared_experts tensors uncompressed. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic rather than routed expert-index traffic.", "notes": "Header-derived bytes are used because the package mixes packed I32 AWQ tensors, BF16 tensors, small F32 tensors, and I64 shape tensors. Routed expert tensors are byte-uniform across all 128 expert indexes and include weight_packed, weight_scale, and weight_shape tensors across MoE layers 1-45." }, "kv_adapter": { "kind": "full_context", "layers": 46, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 46 decoder layers, 8 KV heads, 128 head dimension, and no sliding-window or latent-cache setting. The audited Transformers glm4_moe attention path updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "Glm4MoeForCausalLM profile for ordinary cached text decode. The config records num_nextn_predict_layers 1, but the safetensors index contains no MTP/next-token sidecar tensors." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-glm4-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, scales, shape tensors, and unquantized tensors from safetensors data offsets. AWQ dequantization, activation traffic, router compute, expert compute, and KV writes are outside this memory-side bound.", "notes": "The config records dtype bfloat16 and compressed-tensors pack-quantized int4 weights with group size 32. kv_cache_scheme is null, so this profile charges BF16 K/V cache streams." }, "evidence": [ { "label": "cyankiwi GLM 4.5 Air AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/GLM-4.5-Air-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 982ca6089b239302649a166da300ba83017391ff, the current API records a public MIT text-generation repo with transformers, safetensors, glm4_moe, compressed-tensors, endpoints_compatible, region:us, base_model zai-org/GLM-4.5-Air, and safetensors metadata split across I32, BF16, F32, and I64 tensors. Current downloads were 511138 when audited." }, { "label": "cyankiwi GLM 4.5 Air AWQ config", "url": "https://huggingface.co/cyankiwi/GLM-4.5-Air-AWQ-4bit/raw/982ca6089b239302649a166da300ba83017391ff/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Glm4MoeForCausalLM, glm4_moe, dtype bfloat16, 46 layers, hidden size 4096, intermediate size 10944, MoE intermediate size 1408, 96 attention heads, 8 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, 1 shared expert, first_k_dense_replace 1, 131072 max position embeddings, untied embeddings, vocab size 151552, and compressed-tensors pack-quantized AWQ-style int4 weights with group size 32. The quantization ignore list excludes layer 0, shared experts, and lm_head." }, { "label": "GLM 4.5 Air base config", "url": "https://huggingface.co/zai-org/GLM-4.5-Air/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching architecture, model type, hidden size, MLP sizes, layer count, attention head count, KV head count, head dimension, context length, MoE expert count, experts per token, shared expert count, first_k_dense_replace, dtype, untied embeddings, vocabulary size, and RoPE theta between the quantized config and base config." }, { "label": "cyankiwi GLM 4.5 Air AWQ safetensors headers", "url": "https://huggingface.co/cyankiwi/GLM-4.5-Air-AWQ-4bit/raw/982ca6089b239302649a166da300ba83017391ff/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across all 13 indexed shards using tensor data_offsets. Stored tensors sum to the index total_size, 63.394045248 GB, across 52845 tensors: 52.28199936 GB I32, 11.111743488 GB BF16, 0.00027936 GB I64, and 0.00002304 GB F32. model.embed_tokens.weight is 1.241513984 GB resident-only. Routed expert tensors under model.layers.*.mlp.experts.* total 56.05714944 GB and divide exactly into 128 uniform expert indexes of 0.43794648 GB. Fixed ordinary text traffic, including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head.weight, totals 6.095381824 GB." }, { "label": "Transformers GLM4 MoE implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm4_moe/modeling_glm4_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found Glm4MoeAttention projects key_states and value_states with num_key_value_heads, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx). This supports charging expanded full-context BF16 K/V cache streams for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served quantized config, base config comparison, direct safetensors header byte grouping, and the upstream Transformers glm4_moe runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the artifact as ideal 4-bit MoE weights and did not account for unquantized BF16 shared/fixed tensors, scale tensors, shape tensors, the untied output head, and resident-only input embedding." }, { "id": "cyankiwi--hermes-4-14b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Hermes-4-14B-AWQ-4bit", "title": "cyankiwi Hermes 4 14B AWQ 4-bit", "summary": "Audited memory-side bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Hermes 4 14B.", "model_family": "qwen3-dense-awq", "base_model_proof": { "base_model": "NousResearch/Hermes-4-14B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served AWQ config comparison, Hermes BF16 config, and Qwen3 14B profile comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records NousResearch/Hermes-4-14B as its quantized base model. Manual comparison found matching architecture fields between this AWQ config, the BF16 Hermes config, and the audited Qwen3 14B geometry: Qwen3ForCausalLM, 40 layers, hidden size 5120, intermediate size 17408, 40 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, rope_theta 1000000, tie_word_embeddings false, and vocab size 151936. The AWQ artifact adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "hermes-4-14b-qwen3", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7683072, "swept_params_b": 13.99039488, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 10.54428608, "swept_weight_gb": 8.98846144, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "logical Qwen3/Hermes 14B parameters represented by compressed-tensors safetensors headers", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, compressed-tensors metadata, scales, and lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 weight_packed tensors, BF16 weight_scale tensors, I64 weight_shape tensors, and unquantized BF16 embedding/head/norm tensors. Logical parameter counts follow the audited Qwen3/Hermes 14B architecture: I32 weight_packed tensors are counted as eight 4-bit logical values each, BF16 model weights are counted by element, and weight_scale/weight_shape tensors are storage overhead rather than model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false and sliding_window null, so this profile charges full-context K and V streams for all 40 language layers." }, "notes": "This is a dense Qwen3ForCausalLM text-generation profile. Hermes chat/reasoning/tool-use behavior affects prompting and outputs, not the memory-side decoder adapter." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.713980684258789, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights, BF16 scales, I64 shape tensors, unquantized BF16 tensors, embeddings, output head, and norms from safetensors headers. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, ignore lm_head, and kv_cache_scheme null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "cyankiwi Hermes 4 14B AWQ 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Hermes-4-14B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 947514e4664a053aced581e6e51d5cb9af0623d2, the API records a public Apache-2.0 text-generation repo derived from NousResearch/Hermes-4-14B, with qwen3, compressed-tensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 606230. The API safetensors block reports BF16 1969126400, I32 13212057600, I64 560, and total 3620634160 storage-accounting elements." }, { "label": "cyankiwi Hermes 4 14B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Hermes-4-14B-AWQ-4bit/raw/947514e4664a053aced581e6e51d5cb9af0623d2/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM, hidden size 5120, intermediate size 17408, 40 layers, 40 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, rope_theta 1000000, vocab size 151936, tie_word_embeddings false, use_sliding_window false, sliding_window null, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, ignore lm_head, and kv_cache_scheme null." }, { "label": "NousResearch Hermes 4 14B base config", "url": "https://huggingface.co/NousResearch/Hermes-4-14B/raw/d6ce765c8b83f847357b98254be079afa0c6ca76/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison found no differences in the architecture fields used by this profile between the BF16 Hermes config and the compressed-tensors AWQ config after excluding quantization metadata and dtype labels." }, { "label": "Qwen3 14B audited profile comparison", "url": "https://huggingface.co/Qwen/Qwen3-14B", "source_type": "manual_review", "supports": [ "architecture_compatibility", "logical_parameter_count" ], "notes": "The audited Qwen3 14B BF16 profile records the same served Qwen3ForCausalLM geometry and logical split used here: 14.7683072B resident logical parameters, 13.99039488B swept ordinary-decode logical parameters, and a 0.77791232B resident-only input embedding." }, { "label": "cyankiwi Hermes 4 14B AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Hermes-4-14B-AWQ-4bit/resolve/947514e4664a053aced581e6e51d5cb9af0623d2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 1003 tensors across three shards and records total_size 10544286080 bytes. Direct range-read safetensors headers match the index exactly: 10.54428608 GB total tensor payload bytes, split into I32 6.6060288 GB, BF16 3.9382528 GB, and I64 0.00000448 GB. Stored suffix bytes are weight_packed 6.6060288 GB, weight_scale 0.8257536 GB, weight_shape 0.00000448 GB, and unquantized BF16 weight tensors 3.1124992 GB. model.embed_tokens.weight has shape [151936, 5120] and contributes 1.55582464 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. Swept ordinary text tensors, defined as model.layers plus model.norm.weight plus lm_head.weight and compressed-tensors side tensors, total 8.98846144 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the served compressed-tensors config, BF16 Hermes config comparison, audited Qwen3 14B profile comparison, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, and norms." }, { "id": "cyankiwi--nex-n2-mini-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Nex-N2-mini-AWQ-INT4", "title": "cyankiwi Nex-N2 mini AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ INT4 package of Nex-N2-mini.", "model_family": "nex-n2-mini-qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "nex-agi/Nex-N2-mini", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, BF16 base config comparison, and safetensors header review", "config_compatible": true, "notes": "The compressed-tensors artifact records nex-agi/Nex-N2-mini as its quantized base model. Manual comparison with the current base config found no differences in the checked top-level, text_config, and vision_config geometry fields after excluding quantization_config and dtype packaging. The Nex-N2 card describes Nex-N2-mini as built on Qwen3.5-35B-A3B-Base." }, "architecture": { "canonical_architecture_id": "nex-n2-mini-qwen35moe", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.413055712, "main_resident_weight_gb": 22.502794496, "auxiliary_resident_weight_gb": 1.910261216, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.07274688, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_f16_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding and vision tower tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token; the indexed package has no mtp-named tensors despite mtp_num_hidden_layers 1 in config", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. The compressed-tensors ignore list leaves shared_expert and shared_expert_gate modules uncompressed; shared expert traffic is included in fixed_weight_gb rather than routed expert traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 INT4 payload tensors, F16 scale and zero-point tensors, BF16/F16 ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full_attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The served config preserves the Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Nex-N2-mini text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and speculative MTP decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6576702791798893, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-nex-n2-mini-qwen3.5-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed INT4 I32 tensors, F16 scale and zero-point tensors, unquantized F16/BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, asymmetric quantization with int8 zero-points, F16 dtype packaging, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by API logical safetensors parameters; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Nex-N2-mini AWQ INT4 model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Nex-N2-mini-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 00ff7ea4cd56acf5ed4c0d8144fb6cb68765b446, the API records a public Apache-2.0 repo derived from nex-agi/Nex-N2-mini, with transformers, safetensors, qwen3_5_moe, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 108667. The API safetensors block reports I64: 61440, I32: 33218887680, F16: 3454988928, BF16: 446571248, total: 37120509296 logical storage-accounting elements." }, { "label": "cyankiwi Nex-N2-mini AWQ INT4 config", "url": "https://huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4/raw/00ff7ea4cd56acf5ed4c0d8144fb6cb68765b446/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and asymmetric quantization, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, mtp_num_hidden_layers 1, and kv_cache_scheme null." }, { "label": "Nex-N2-mini base config comparison", "url": "https://huggingface.co/nex-agi/Nex-N2-mini/raw/ca218dcb1fbe05f84d1807d180cb5d9bcb1c5c93/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited top-level, text_config, and vision_config geometry fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config, dtype packaging, and Transformers bookkeeping fields." }, { "label": "cyankiwi Nex-N2-mini AWQ INT4 safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4/resolve/00ff7ea4cd56acf5ed4c0d8144fb6cb68765b446/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all five indexed shards. Stored tensor payloads sum to 24.413055712 GB across 123826 tensors: I32 16.609443840 GB, F16 6.909977856 GB, BF16 0.893142496 GB, and I64 0.000491520 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 22.502794496 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1.910261216 GB. The indexed headers contain no mtp-named tensors. Routed expert tensors sum to 18.623201280 GB and divide exactly into 256 uniform expert indexes of 0.072746880 GB. Fixed ordinary text traffic sums to 3.879593216 GB. Linked shard objects sum to 24.430464432 GB, leaving 0.017408720 GB of safetensors header/container overhead outside tensor payloads." }, { "label": "Nex-N2 model card", "url": "https://huggingface.co/nex-agi/Nex-N2-mini", "source_type": "model_card", "supports": [ "base_model_proof", "model_family" ], "notes": "The Nex-N2 card states that Nex-N2-mini is built on Qwen3.5-35B-A3B-Base and recommends Nex-series deployment with the customized SGLang fork plus qwen3 reasoning/tool parsers." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card, pinned compressed-tensors config, current BF16 base config comparison, direct safetensors index resolution, range-read safetensors shard headers, linked-object HEAD checks, and the existing Transformers qwen3_5 runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared expert, and charged all 40 layers as full-context KV instead of the 10 full-attention layers plus fixed DeltaNet state used by Qwen3.5." }, { "id": "cyankiwi--qwen3-30b-a3b-instruct-2507-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit", "title": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ 4-bit", "summary": "Audited memory-side bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3 30B A3B Instruct 2507.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B-Instruct-2507", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, BF16 base config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3-30B-A3B-Instruct-2507 as its quantized base model. Manual comparison found matching checked architecture fields between the public BF16 base config and this artifact; the artifact adds quantization_config and uses dtype instead of torch_dtype while preserving the Qwen3MoeForCausalLM geometry." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.087607296, "main_resident_weight_gb": 17.46527744, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.157528576, "routed_expert_weight_gb": 0.127404288, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 shape tensors. Routed expert tensors are byte-uniform across all 128 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors. The router gate tensors are uncompressed BF16 and included in fixed_weight_gb." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served compressed-tensors config and direct safetensors header grouping. The model card notes native 262144-token context; this profile uses the native served max_position_embeddings value." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5924123756067358, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, and long-context attention kernels are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by the audited BF16 logical model parameter count; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 876f886e383033237a04d8634b5af1cde0818290, the API records a public Apache-2.0 text-generation repo derived from Qwen/Qwen3-30B-A3B-Instruct-2507, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 534049. The API safetensors block reports I32 29896998912 logical 4-bit values, BF16 1569404928 elements, and I64 37248 shape elements; the safetensors index records total_parameters 5306567040 storage-accounting elements and total_size 18087607296 bytes." }, { "label": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ model card", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The card identifies the base model as Qwen/Qwen3-30B-A3B-Instruct-2507 and describes the non-thinking 2507 model as 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, and 262144 native context." }, { "label": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 dtype, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, vocab_size 151936, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, MSE observer, and kv_cache_scheme null." }, { "label": "Qwen3 30B A3B Instruct 2507 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in checked architecture fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config and dtype key spelling." }, { "label": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four indexed shards. Stored tensor payloads sum to 18.087607296 GB across 56115 tensors: BF16 3.138809856 GB, I32 14.948499456 GB, and I64 0.000297984 GB. Ordinary text resident tensors, defined as all tensors except model.embed_tokens.weight, sum to 17.465277440 GB. model.embed_tokens.weight is resident-only and contributes 0.622329856 GB. Routed expert tensors sum to 16.307748864 GB and divide exactly into 128 uniform expert indexes of 0.127404288 GB. Fixed ordinary text traffic, including attention, routers, norms, and lm_head.weight, sums to 1.157528576 GB." }, { "label": "cyankiwi Qwen3 30B A3B Instruct 2507 AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring lm_head, model.embed_tokens, input/post-attention layer norms, model.norm, and router gate modules. The served config and safetensors headers remain authoritative for exact stored tensor bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, BF16 base config comparison, quantization recipe, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the generated ideal 4-bit MoE estimate, which undercounted compressed-tensors side tensors and unquantized attention, router, output, norm, and embedding tensors." }, { "id": "cyankiwi--qwen3-30b-a3b-thinking-2507-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit", "title": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ 4-bit", "summary": "Audited memory-side bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3 30B A3B Thinking 2507.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B-Thinking-2507", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, and direct base-config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3-30B-A3B-Thinking-2507 as its quantized base model. Manual comparison found matching checked architecture fields between the public BF16 base config and this artifact; the artifact adds quantization_config and uses dtype instead of torch_dtype while preserving the Qwen3MoeForCausalLM geometry." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.087607296, "main_resident_weight_gb": 17.46527744, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.157528576, "routed_expert_weight_gb": 0.127404288, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 shape tensors. Routed expert tensors are byte-uniform across all 128 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors. The router gate tensors are uncompressed BF16 and included in fixed_weight_gb." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served compressed-tensors config and direct safetensors header grouping. The model card notes native 262144-token context and optional ultra-long context serving up to about 1M tokens with MInference and Dual Chunk Attention; this profile uses the native served max_position_embeddings value." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5924123756067358, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, and sparse long-context attention kernels are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by the audited BF16 logical model parameter count; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 4f1525a21bef5cbd32587a1f392ce75b281b633e, the API records a public Apache-2.0 text-generation repo derived from Qwen/Qwen3-30B-A3B-Thinking-2507, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 420409. The API safetensors block reports I32 29896998912 logical 4-bit values, BF16 1569404928 elements, and I64 37248 shape elements; the safetensors index records total_parameters 5306567040 storage-accounting elements and total_size 18087607296 bytes." }, { "label": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ model card", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The card identifies the base model as Qwen/Qwen3-30B-A3B-Thinking-2507 and states 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, and 262144 native context. It also describes the model as thinking-only and documents optional 1M-token serving with MInference and Dual Chunk Attention." }, { "label": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 dtype, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, vocab_size 151936, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, MSE observer, and kv_cache_scheme null." }, { "label": "Qwen3 30B A3B Thinking 2507 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in checked architecture fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config and dtype key spelling." }, { "label": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four indexed shards. Stored tensor payloads sum to 18.087607296 GB across 56115 tensors: BF16 3.138809856 GB, I32 14.948499456 GB, and I64 0.000297984 GB. Ordinary text resident tensors, defined as all tensors except model.embed_tokens.weight, sum to 17.465277440 GB. model.embed_tokens.weight is resident-only and contributes 0.622329856 GB. Routed expert tensors sum to 16.307748864 GB and divide exactly into 128 uniform expert indexes of 0.127404288 GB. Fixed ordinary text traffic, including attention, routers, norms, and lm_head.weight, sums to 1.157528576 GB." }, { "label": "cyankiwi Qwen3 30B A3B Thinking 2507 AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring lm_head, model.embed_tokens, input/post-attention layer norms, model.norm, and router gate modules. The served config and safetensors headers remain authoritative for exact stored tensor bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, BF16 base config comparison, quantization recipe, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the generated ideal 4-bit MoE estimate, which undercounted compressed-tensors side tensors and unquantized attention, router, output, norm, and embedding tensors." }, { "id": "cyankiwi--qwen3-5-122b-a10b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit", "title": "cyankiwi Qwen3.5 122B A10B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 122B A10B.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.5-122B-A10B as its quantized base model. Manual comparison found no differences in the checked top-level, text, vision, MoE, attention, and DeltaNet state geometry fields between the base BF16 config and this artifact; the target adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 80.00229424, "main_resident_weight_gb": 75.999547392, "auxiliary_resident_weight_gb": 4.002746848, "fixed_weight_gb": 10.76914176, "routed_expert_weight_gb": 0.254806272, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "base logical Qwen3.5 122B A10B parameters with direct compressed-tensors safetensors stored-byte totals", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024. The compressed-tensors recipe and served quantization ignore list leave shared_expert and shared_expert_gate modules uncompressed; shared expert traffic is included in fixed_weight_gb rather than routed expert traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 256 expert indexes, including weight, weight_scale, and weight_shape side tensors. The HF API storage-accounting total_size is lower than the direct header payload sum for this artifact, so exact header-derived byte fields drive production bounds." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with every fourth layer using full_attention, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6395757827871972, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 text dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by the base logical 125.086497008B parameter identity; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3.5 122B A10B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 2e7164b035cdf708f05a1c5ce8f532ec39096d40, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-122B-A10B, with compressed-tensors, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 349099. The API safetensors block reports I64: 75264, I32: 118380036096, BF16: 10405837040, and total: 24266876928 storage-accounting elements." }, { "label": "cyankiwi Qwen3.5 122B A10B AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit/raw/2e7164b035cdf708f05a1c5ce8f532ec39096d40/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, 48 text layers, full_attention_interval 4, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, one MTP layer, and kv_cache_scheme null." }, { "label": "Qwen3.5 122B A10B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited top-level, text_config, and vision_config geometry fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config, dtype packaging, and Transformers bookkeeping fields." }, { "label": "cyankiwi Qwen3.5 122B A10B AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit/resolve/2e7164b035cdf708f05a1c5ce8f532ec39096d40/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 16 indexed shards. Stored tensor payloads sum to 80.002294240 GB across 113981 tensors: BF16 20.811674080 GB, I32 59.190018048 GB, and I64 0.000602112 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 75.999547392 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 4.002746848 GB. Routed expert tensors sum to 65.230405632 GB and divide exactly into 256 uniform expert indexes of 0.254806272 GB. Fixed ordinary text traffic sums to 10.769141760 GB. The indexed metadata reports total_size 77.525225472 GB, but direct shard header payloads are used because they are the audited byte source for every tensor." }, { "label": "cyankiwi Qwen3.5 122B A10B AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit/raw/2e7164b035cdf708f05a1c5ce8f532ec39096d40/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention regexes, shared experts, shared expert gates, MoE gates, self-attention modules, visual modules, MTP modules, and lm_head. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, BF16 base config comparison, quantization recipe, direct safetensors index resolution, range-read safetensors shard headers, and the existing Transformers qwen3_5 runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared expert, used all layers as full-context KV, and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, visual, MTP, linear-attention, attention, and shared-expert modules." }, { "id": "cyankiwi--qwen3-5-27b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-27B-AWQ-4bit", "title": "cyankiwi Qwen3.5 27B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 27B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.5-27B as its quantized base model. Manual comparison found matching checked top-level, text, and vision architecture fields: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds compressed-tensors quantization metadata while preserving BF16 text dtype." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 20.057751376, "swept_weight_gb": 16.279196928, "auxiliary_resident_weight_gb": 3.778554448, "resident_parameter_scope": "base logical Qwen3.5 27B parameters with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 weight_packed tensors, BF16 weight_scale tensors, BF16 embeddings/lm_head, BF16 visual/MTP tensors, and tiny I64 shape tensors. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes. The HF API storage-accounting total is 28.553279006B elements because compressed-tensors scale and shape side tensors are counted in the safetensors parameter summary." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7219841762869511, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights, BF16 scales, unquantized BF16 tensors, and tiny I64 shape tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 text dtype, and kv_cache_scheme null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "cyankiwi Qwen3.5 27B AWQ 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-27B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 65497df9ca706cd6aa9cc8374fc259f39e78672d, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-27B, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 557518. The API safetensors block reports BF16: 3854070512, I32: 24699207680, I64: 814, total: 28553279006 storage-accounting elements." }, { "label": "cyankiwi Qwen3.5 27B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit/raw/65497df9ca706cd6aa9cc8374fc259f39e78672d/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, text_config dtype bfloat16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, MTP settings, and kv_cache_scheme null." }, { "label": "Qwen3.5 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked top-level, text_config, layer_types, and vision_config geometry fields between the current base BF16 repo and this compressed-tensors artifact. Both configs record BF16 text dtype and the same hybrid full-attention/linear-attention layer pattern." }, { "label": "cyankiwi Qwen3.5 27B AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit/resolve/65497df9ca706cd6aa9cc8374fc259f39e78672d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 2013 tensors across four shards and has no total_size metadata field. Direct range-read safetensors headers sum to 20.057751376 GB: BF16 7.708141024 GB, I32 12.349603840 GB, and I64 0.000006512 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 16.279196928 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.778554448 GB. Header buckets are language MLP 9.625930752 GB, language linear-attention tensors 3.165412608 GB, lm_head 2.542796800 GB, input embedding 2.542796800 GB, self-attention tensors 0.943735808 GB, visual 0.921460192 GB, MTP 0.314297456 GB, and language layer/norm tensors 0.001320960 GB." }, { "label": "cyankiwi Qwen3.5 27B AWQ 4-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit/raw/65497df9ca706cd6aa9cc8374fc259f39e78672d/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "quantization" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention regexes, visual modules, MTP modules, and lm_head. The served config/header remains authoritative for exact stored tensor bytes because the final artifact contains a mix of packed I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, recipe, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, MTP, visual, and selected language tensors." }, { "id": "cyankiwi--qwen3-5-2b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-2B-AWQ-4bit", "title": "cyankiwi Qwen3.5 2B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 2B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-2B", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The repo metadata records Qwen/Qwen3.5-2B as the quantized base model. Manual comparison found no differences in the checked architecture fields between the AWQ config and the base Qwen3.5 2B config: Qwen3_5ForConditionalGeneration, tied embeddings, 24 text layers, every fourth layer using full attention, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, resident vision tower, one MTP layer, and 262144 max position embeddings. The artifact adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.274069824, "swept_params_b": 1.881825088, "auxiliary_resident_params_b": 0.392244736, "resident_weight_gb": 2.501190736, "swept_weight_gb": 1.792067552, "auxiliary_resident_weight_gb": 0.709123184, "resident_parameter_scope": "logical Qwen3.5 2B parameter split with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model safetensors headers including the tied embed_tokens output projection", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 tensors, BF16 ignored modules, BF16 embeddings, BF16 visual tensors, and BF16/I32 MTP tensors. Logical parameter counts match the audited BF16/F32 Qwen3.5 2B profile. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 24 layers with every fourth layer marked full_attention, giving 6 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. AWQ weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.0998742033349278, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 tensors, BF16 ignored text modules, BF16 embeddings, BF16 visual tensors, and BF16/I32 MTP tensors. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, text_config dtype bfloat16, and no KV cache quantization scheme. KV cache and DeltaNet state are therefore charged at two bytes per scalar for BF16 plus F32 recurrent state." }, "evidence": [ { "label": "cyankiwi Qwen3.5 2B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-2B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA e27b458e4c295748030ed367efaee3b2f2ccf32f, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-2B, with transformers, safetensors, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 465061. The API safetensors block reports I64: 314, I32: 1423966208, BF16: 894602560, total: 2318569082." }, { "label": "cyankiwi Qwen3.5 2B AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-2B-AWQ-4bit/raw/e27b458e4c295748030ed367efaee3b2f2ccf32f/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, no KV cache quantization scheme, bfloat16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and one MTP layer." }, { "label": "Qwen3.5 2B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in 27 checked architecture fields between the current base config and the AWQ artifact after excluding compressed-tensors quantization metadata." }, { "label": "cyankiwi Qwen3.5 2B AWQ safetensors index and shard header", "url": "https://huggingface.co/cyankiwi/Qwen3.5-2B-AWQ-4bit/resolve/e27b458e4c295748030ed367efaee3b2f2ccf32f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 946 tensors into one shard. Direct range-read safetensors headers found tensor payloads totaling 2.501190736 GB: 1.789205120 GB BF16, 0.711983104 GB I32, and 0.000002512 GB I64. The ordinary text swept subset, defined as all model.language_model tensors including the tied output embedding, totals 1.792067552 GB. The resident-only subset, defined as model.visual plus top-level mtp tensors, totals 0.709123184 GB. Header buckets are language_embed 1.017118720 GB, other language_model tensors 0.774948832 GB, visual 0.662833152 GB, and MTP 0.046290032 GB. The linked safetensors file is 2.501307856 GB, leaving 117120 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, safetensors index, direct range-read safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and missed BF16 ignored tensors, resident visual/MTP tensors, exact AWQ side-tensor storage, and fixed DeltaNet runtime state." }, { "id": "cyankiwi--qwen3-5-35b-a3b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit", "title": "cyankiwi Qwen3.5 35B A3B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 35B A3B.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.5-35B-A3B as its quantized base model. Manual comparison found no differences in the checked top-level, text, vision, MoE, attention, and DeltaNet state geometry fields between the base BF16 config and this artifact; the target adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.441405152, "main_resident_weight_gb": 21.999478016, "auxiliary_resident_weight_gb": 2.441927136, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.0707808, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. The compressed-tensors recipe and served quantization ignore list leave shared_expert and shared_expert_gate modules uncompressed; shared expert traffic is included in fixed_weight_gb rather than routed expert traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 256 expert indexes, including weight, weight_scale, and weight_shape side tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full_attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6608699351139912, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 text dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by API logical safetensors parameters; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3.5 35B A3B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA affb52d39eeb1c4d702372be794c7c2b36c32d47, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-35B-A3B, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 441851. The API safetensors block reports I64: 62976, I32: 33017561088, BF16: 3966060400, total: 36983684464 storage-accounting elements." }, { "label": "cyankiwi Qwen3.5 35B A3B AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit/raw/affb52d39eeb1c4d702372be794c7c2b36c32d47/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, one MTP layer, and kv_cache_scheme null." }, { "label": "Qwen3.5 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/59d61f3ce65a6d9863b86d2e96597125219dc754/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited top-level, text_config, and vision_config geometry fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config, dtype packaging, and Transformers bookkeeping fields." }, { "label": "cyankiwi Qwen3.5 35B A3B AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit/resolve/affb52d39eeb1c4d702372be794c7c2b36c32d47/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all five indexed shards. Stored tensor payloads sum to 24.441405152 GB across 95427 tensors: BF16 7.932120800 GB, I32 16.508780544 GB, and I64 0.000503808 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.999478016 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 2.441927136 GB. Routed expert tensors sum to 18.119884800 GB and divide exactly into 256 uniform expert indexes of 0.070780800 GB. Fixed ordinary text traffic sums to 3.879593216 GB." }, { "label": "cyankiwi Qwen3.5 35B A3B AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit/raw/affb52d39eeb1c4d702372be794c7c2b36c32d47/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention regexes, shared experts, attention modules, visual modules, MTP modules, gate modules, and lm_head. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, BF16 base config comparison, quantization recipe, direct safetensors index resolution, range-read safetensors shard headers, and the existing Transformers qwen3_5 runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared expert, and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, visual, MTP, linear-attention, attention, and shared-expert modules." }, { "id": "cyankiwi--qwen3-5-4b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-4B-AWQ-4bit", "title": "cyankiwi Qwen3.5 4B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 4B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-4B", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The repo metadata records Qwen/Qwen3.5-4B as the quantized base model. Manual comparison found no differences in the checked architecture fields between the AWQ config and the base Qwen3.5 4B config: Qwen3_5ForConditionalGeneration, tied embeddings, 32 text layers, every fourth layer using full attention, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, one MTP layer, and 262144 max position embeddings. The artifact adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.659865088, "swept_params_b": 4.205751296, "auxiliary_resident_params_b": 0.454113792, "resident_weight_gb": 4.040317168, "swept_weight_gb": 3.286590592, "auxiliary_resident_weight_gb": 0.753726576, "resident_parameter_scope": "logical Qwen3.5 4B parameter split with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model safetensors headers including the tied embed_tokens output projection", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 tensors, BF16 ignored modules, BF16 embeddings, BF16 visual tensors, and BF16/I32 MTP tensors. Logical parameter counts match the audited BF16/F32 Qwen3.5 4B profile. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. AWQ weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.867045953412804, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 tensors, BF16 ignored text modules, BF16 embeddings, BF16 visual tensors, and BF16/I32 MTP tensors. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 text config, and no KV cache quantization scheme. KV cache and DeltaNet state are therefore charged at two bytes per scalar for BF16 plus F32 recurrent state." }, "evidence": [ { "label": "cyankiwi Qwen3.5 4B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-4B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA b878ff99819d1edba0d167876eaa2e4faef5d9d7, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-4B, with transformers, safetensors, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 363086. The API safetensors block reports I64: 414, I32: 3672637440, BF16: 1101997568, total: 4774635422." }, { "label": "cyankiwi Qwen3.5 4B AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit/raw/b878ff99819d1edba0d167876eaa2e4faef5d9d7/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, no KV cache quantization scheme, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and one MTP layer." }, { "label": "Qwen3.5 4B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the checked top-level and text architecture fields between the current base config and the AWQ artifact after excluding compressed-tensors quantization metadata." }, { "label": "cyankiwi Qwen3.5 4B AWQ safetensors index and shard header", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit/resolve/b878ff99819d1edba0d167876eaa2e4faef5d9d7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 1152 tensors into one shard. Direct range-read safetensors headers found tensor payloads totaling 4.040317168 GB: 2.203995136 GB BF16, 1.836318720 GB I32, and 0.000003312 GB I64. The ordinary text swept subset, defined as all model.language_model tensors including the tied output embedding, totals 3.286590592 GB. The resident-only subset, defined as model.visual plus top-level mtp tensors, totals 0.753726576 GB. Header buckets are language_embed 1.271398400 GB, language_mlp 1.274021376 GB, visual 0.667028480 GB, language_linear_attn 0.575678592 GB, language_self_attn 0.165159424 GB, MTP 0.086698096 GB, language_layer_norm 0.000327680 GB, and language_final_norm 0.000005120 GB. The linked safetensors file is 4.040461440 GB, leaving 144272 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "cyankiwi Qwen3.5 4B AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit/raw/b878ff99819d1edba0d167876eaa2e4faef5d9d7/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "quantization" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention regexes, visual modules, MTP modules, and lm_head. The served config/header remains authoritative for exact stored tensor bytes because the final artifact contains a mix of packed I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, recipe, safetensors index, direct range-read safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and missed BF16 ignored tensors, resident visual/MTP tensors, exact AWQ side-tensor storage, and fixed DeltaNet runtime state." }, { "id": "cyankiwi--qwen3-5-9b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.5-9B-AWQ-4bit", "title": "cyankiwi Qwen3.5 9B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.5-9B as its quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds compressed-tensors quantization metadata while preserving BF16 text dtype." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 9.068440272, "swept_weight_gb": 5.937066112, "auxiliary_resident_weight_gb": 3.13137416, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 weight_packed tensors, BF16 weight_scale tensors, BF16 embeddings/lm_head, BF16 visual/MTP tensors, and tiny I64 shape tensors. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 9.653104368B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.9394325313690631, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights, BF16 scales, unquantized BF16 tensors, and tiny I64 shape tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, and BF16 text dtype. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "cyankiwi Qwen3.5 9B AWQ 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-9B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 73536aa464f9a93c550aa5a916f0113a08b2f384, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 758044. The API safetensors block reports BF16: 2753736432, I32: 7121928192, I64: 414, total: 9875665038 storage-accounting elements." }, { "label": "cyankiwi Qwen3.5 9B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit/raw/73536aa464f9a93c550aa5a916f0113a08b2f384/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, text_config dtype bfloat16, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings. The quantization config records kv_cache_scheme null." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching top-level identity fields and text geometry fields between the base BF16 repo and this compressed-tensors artifact. Both configs record BF16 text dtype and the same hybrid full-attention/linear-attention layer pattern." }, { "label": "cyankiwi Qwen3.5 9B AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit/resolve/73536aa464f9a93c550aa5a916f0113a08b2f384/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 1189 tensors across three shards. Direct range-read safetensors headers sum to 9.068440272 GB: BF16 5.507472864 GB, I32 3.560964096 GB, and I64 0.000003312 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 5.937066112 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.131374160 GB. Header buckets are language_mlp 2.717910528 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, language_linear_attn 0.920135808 GB, visual 0.912020960 GB, language_self_attn 0.264249856 GB, MTP 0.185115760 GB, language layer/norm 0.000532480 GB, and language final norm 0.000008192 GB." }, { "label": "cyankiwi Qwen3.5 9B AWQ 4-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit/raw/73536aa464f9a93c550aa5a916f0113a08b2f384/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "quantization" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention regexes, visual modules, MTP modules, and lm_head. The served config/header remains authoritative for exact stored tensor bytes because the final artifact contains a mix of packed I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, recipe, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, MTP, visual, and selected language tensors." }, { "id": "cyankiwi--qwen3-6-27b-awq-bf16-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4", "title": "Cyankiwi Qwen3.6 27B AWQ BF16 INT4", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors BF16/INT4 package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.6-27B as its quantized base model. Manual comparison found matching checked top-level and text architecture fields against both Qwen/Qwen3.6-27B and cyankiwi/Qwen3.6-27B-AWQ-INT4: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds compressed-tensors quantization metadata and records text dtype float16." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 28.315858512, "swept_weight_gb": 24.531487744, "auxiliary_resident_weight_gb": 3.784370768, "resident_parameter_scope": "logical Qwen3.6 27B model parameter split with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 tensors, F16 scales/ignored modules/embeddings/lm_head, and BF16 visual/MTP tensors. Logical parameter counts follow the audited Qwen3.6 BF16 profile; storage metadata tensors are byte traffic, not separate model parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.0192369723011852, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-bf16-int4-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights/metadata, F16 scales and ignored text tensors, F16 embeddings/lm_head, and unquantized BF16 visual/MTP tensors. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, asymmetric zero points, text_config dtype float16, and 400 ignored module names including the vision tower, linear-attention projections, mtp.fc, and lm_head. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Cyankiwi Qwen3.6 27B AWQ BF16 INT4 model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA a3d9607a84817d20ed404d0d12765e70c8a2026f, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-27B, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 570671. The API safetensors block reports BF16: 524817648, F16: 8692719104, I32: 19761561600, I64: 526, total: 28979098878 storage-accounting elements." }, { "label": "Cyankiwi Qwen3.6 27B AWQ BF16 INT4 config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4/raw/a3d9607a84817d20ed404d0d12765e70c8a2026f/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and asymmetric zero points, text_config dtype float16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in checked top-level and text_config geometry fields between the current base BF16 repo, this BF16/INT4 artifact, and the existing cyankiwi AWQ-INT4 artifact. The artifacts add quantization_config and dtype labels but preserve the architecture fields used by this profile." }, { "label": "Cyankiwi Qwen3.6 27B AWQ BF16 INT4 safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4/resolve/a3d9607a84817d20ed404d0d12765e70c8a2026f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 1988 tensors across six shards. Its metadata total_size is 320113776 bytes, but direct range-read safetensors headers sum to 28.315858512 GB, so this profile uses header-derived bytes. Header dtype bytes are BF16 1.049635296 GB, F16 17.385438208 GB, I32 9.880780800 GB, and I64 0.000004208 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 24.531487744 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.784370768 GB. Header buckets are language_other 21.988690944 GB, lm_head 2.542796800 GB, input embedding 2.542796800 GB, visual 0.921460192 GB, and MTP 0.320113776 GB. Linked-object HEAD checks resolved all six shards to 28.316110952 GB, leaving 252440 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, comparison against the existing AWQ-INT4 profile, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted the F16 ignored text tensors, zero points/scales, embeddings, output head, MTP, visual, and selected linear-attention tensors." }, { "id": "cyankiwi--qwen3-6-27b-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.6-27B-AWQ-INT4", "title": "Cyankiwi Qwen3.6 27B AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors INT4 package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3.6-27B as its quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The package adds compressed-tensors quantization metadata and changes text dtype to float16." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 20.443676496, "swept_weight_gb": 16.659305728, "auxiliary_resident_weight_gb": 3.784370768, "resident_parameter_scope": "logical Qwen3.6 27B model parameter split with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 weight_packed/weight_zero_point tensors, F16 weight_scale tensors, F16 embeddings/lm_head, and BF16 visual/MTP tensors. Logical parameter counts follow the audited Qwen3.6 BF16 profile; storage metadata tensors are byte traffic, not separate model parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7358756551794972, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-int4-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights and zero points, F16 scales, F16 embeddings/lm_head, and unquantized BF16 visual/MTP tensors. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, asymmetric zero points, and text_config dtype float16. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Cyankiwi Qwen3.6 27B AWQ INT4 model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-27B-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA c9b937c5466c5c0575fc15edd1f8c516cb1e62fd, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-27B, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 953536. The API safetensors block reports BF16: 524817648, F16: 3329252864, I32: 25471057920, I64: 814, total: 29325129246 storage-accounting elements." }, { "label": "Cyankiwi Qwen3.6 27B AWQ INT4 config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-INT4/raw/c9b937c5466c5c0575fc15edd1f8c516cb1e62fd/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and asymmetric zero points, text_config dtype float16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in checked top-level and text_config geometry fields between the current base BF16 repo and this compressed-tensors artifact. The artifact adds quantization_config and dtype labels but preserves the architecture fields used by this profile." }, { "label": "Cyankiwi Qwen3.6 27B AWQ INT4 safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-INT4/resolve/c9b937c5466c5c0575fc15edd1f8c516cb1e62fd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 2420 tensors across four shards. Its metadata total_size is 320113776 bytes, but direct range-read safetensors headers sum to 20.443676496 GB, so this profile uses header-derived bytes. Header dtype bytes are BF16 1.049635296 GB, F16 6.658505728 GB, I32 12.73552896 GB, and I64 0.000006512 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 16.659305728 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.784370768 GB. Header buckets are language_other 14.116508928 GB, lm_head 2.5427968 GB, input embedding 2.5427968 GB, visual 0.921460192 GB, and MTP 0.320113776 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current base config comparison, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted compressed-tensors zero points/scales plus unquantized embeddings, output head, MTP, visual, and selected linear-attention tensors." }, { "id": "cyankiwi--qwen3-6-35b-a3b-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit", "title": "cyankiwi Qwen3.6 35B A3B AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi AWQ 4-bit package of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ artifact records Qwen/Qwen3.6-35B-A3B as its base model and preserves the audited Qwen3.6 text and vision geometry: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 experts per token, 1 shared expert, and one MTP layer." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.956800736, "main_resident_weight_gb": 22.502302976, "auxiliary_resident_weight_gb": 2.45449776, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.07274496, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_f16_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512, and the quantization config leaves shared_expert and shared_expert_gate modules unconverted. Shared expert traffic is therefore included in fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived bytes are used because the package stores AWQ I32 packed tensors, F16 side tensors, and BF16 tensors. Routed expert tensors are byte-uniform across all 256 expert indexes, including qweight, qzeros, and scales tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. AWQ weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, qzeros, scales, and unquantized tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group size 32 and zero points. KV and DeltaNet state traffic are charged from the preserved base Qwen3.6 runtime geometry." }, "evidence": [ { "label": "cyankiwi Qwen3.6 35B A3B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 87df542ee375fe6da387b555f2ed89e937063023, the API records an Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-35B-A3B with AWQ/4-bit tags and safetensors parameters split across I32: 33017561088, F16: 2448355968, and BF16: 485905648 tensors. The model card identifies cyankiwi AWQ version 26.05.01 with STEM and Agentic calibration." }, { "label": "cyankiwi Qwen3.6 35B A3B AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit/raw/87df542ee375fe6da387b555f2ed89e937063023/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, AWQ quantization with group size 32, zero points, GEMM version, a 40-layer text config, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, and one MTP layer." }, { "label": "Qwen3.6 35B A3B base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The base card states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo and this AWQ artifact; the AWQ artifact adds quantization_config while preserving the base architecture." }, { "label": "cyankiwi Qwen3.6 35B A3B AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit/resolve/87df542ee375fe6da387b555f2ed89e937063023/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The index records total_size 24956800736 bytes across six shards. Safetensors headers were range-read across all six shards and found 95,427 tensors totaling 24.956800736 GB: 17.024679936 GB I32 tensors, 6.960309504 GB F16 tensors, and 0.971811296 GB BF16 tensors. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 22.502302976 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 2.45449776 GB. Main routed expert tensors sum to 18.62270976 GB, or 0.07274496 GB per expert index. Fixed ordinary text traffic sums to 3.879593216 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from model card/API metadata, served config, base config comparison, quantization config, safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate for this AWQ repo, including the catalog row's missing shared-expert count and understated active traffic." }, { "id": "cyankiwi--qwen3-coder-30b-a3b-instruct-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit", "title": "Cyankiwi Qwen3-Coder 30B A3B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3-Coder 30B A3B Instruct.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, official BF16 base config comparison, recipe.yaml, and safetensors header review", "config_compatible": false, "notes": "The repo metadata records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the core ordinary-decode topology used by this profile: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, 262144 max positions, and disabled sliding-window attention. Manual comparison against the audited official BF16 base config found served-config differences in intermediate_size, max_window_layers, router_aux_loss_coef, and quantization metadata, so this profile uses the cyankiwi served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-awq-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.087607296, "main_resident_weight_gb": 17.46527744, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.157528576, "routed_expert_weight_gb": 0.127404288, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_bf16_i64", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are ignored by quantization and included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived bytes are used because the AWQ package stores packed I32 int4 tensors plus BF16 side tensors and tiny I64 shape tensors. Routed expert tensors are stored as per-expert down_proj, gate_proj, and up_proj weight_packed, weight_scale, and weight_shape tensors across all 48 layers; all 128 expert indexes are byte-uniform." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, and the compressed-tensors quantization config records kv_cache_scheme null, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM AWQ compressed-tensors profile using the served cyankiwi config, recipe, and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, dequantization, activation traffic, and expert-parallel communication are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llm-compressor-compressed-tensors-awq-groupsize32-qwen3-coder-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors payload bytes: packed I32 int4 weights, BF16 scales and ignored modules, and tiny I64 shape side tensors from safetensors headers. Dequantization, kernel choice, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group size 32, symmetric group quantization, and kv_cache_scheme null. Runtime KV traffic is charged as BF16 because the target config does not define a KV-cache quantization scheme." }, "evidence": [ { "label": "Cyankiwi Qwen3-Coder 30B A3B AWQ 4-bit API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_moe, conversational, endpoints_compatible, compressed-tensors, region:us, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 245223. The API safetensors summary records I64 37248, I32 29896998912, BF16 1569404928, and total 5306567040, so exact memory figures in this profile come from direct safetensors header reads." }, { "label": "Cyankiwi Qwen3-Coder 30B A3B AWQ 4-bit model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/raw/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base and the compressed-tensors AWQ 4-bit package." }, { "label": "Cyankiwi Qwen3-Coder 30B A3B AWQ 4-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/raw/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/recipe.yaml", "source_type": "config", "supports": [ "serving", "weight_format", "fixed_weight_gb" ], "notes": "The pinned recipe uses AWQModifier targets [Linear], scheme 4-bit integer group quantization with group_size 32, symmetric weights, observer mse, duo_scaling true, and ignore entries for lm_head, model.embed_tokens, input/post attention norms, model.norm, and mlp.gate. This supports treating embeddings as resident-only and lm_head/router/norm tensors as fixed ordinary-decode traffic." }, { "label": "Cyankiwi Qwen3-Coder 30B A3B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/raw/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, qkv_bias false, use_qk_norm true, and compressed-tensors pack-quantized INT4 weights with group size 32, symmetric group strategy, kv_cache_scheme null, and quantization_status compressed." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked the audited architecture fields. The cyankiwi config matches the official BF16 base on the MoE routing and KV-bearing decode fields used by this profile, including layer count, heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, max_window_layers 28 vs 48, router_aux_loss_coef 0.0 vs 0.001, shared_expert_intermediate_size, explicit qkv fields, and quantization metadata, so config_compatible is false and fixed_weight_gb is derived from the cyankiwi tensor headers." }, { "label": "Cyankiwi Qwen3-Coder 30B A3B AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/resolve/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 18087607296 bytes across four shards. Direct range-read safetensors headers found 56115 tensors with payload bytes exactly matching the index total: 18.087607296 GB, split into 14.948499456 GB I32 packed tensors, 3.138809856 GB BF16 tensors, and 0.000297984 GB I64 weight_shape tensors. Linked shard sizes total 18.094507352 GB, leaving 0.006900056 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately and remains in fixed decode traffic. Routed expert tensors sum to 16.307748864 GB and divide exactly into 128 uniform expert groups of 0.127404288 GB. Non-expert fixed decode tensors including lm_head.weight, router/gate tensors, norms, attention, and dense side tensors sum to 1.157528576 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, pinned model card, pinned recipe, pinned served compressed-tensors config, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated metadata estimate for this AWQ repo. It uses exact stored compressed-tensors payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "cyankiwi--qwen3-coder-next-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-Coder-Next-AWQ-4bit", "title": "Cyankiwi Qwen3-Coder-Next AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3-Coder-Next.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, pinned config comparison, compressed-tensors recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The API metadata records Qwen/Qwen3-Coder-Next as the quantized base model. Manual comparison found matching memory-relevant architecture fields between the cyankiwi config and the audited official BF16 base config: architecture, model type, layer count, full_attention_interval, hidden size, expert geometry, attention heads, KV heads, head dimensions, linear-attention state geometry, max positions, vocabulary size, and untied embeddings. The cyankiwi config adds compressed-tensors AWQ quantization metadata while retaining dtype bfloat16." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 48.217683456, "main_resident_weight_gb": 47.5953536, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.08493696, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "resolved safetensors index total_size and direct shard-header stored bytes for the compressed-tensors AWQ package", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed BF16 language tensors plus expected distinct routed expert tensor groups", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. The compressed-tensors recipe excludes shared expert, shared expert gate, and router gate tensors from quantization, so they remain BF16 and are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "The AWQ package only quantizes routed expert Linear tensors. The recipe excludes embeddings, lm_head, norms, full attention, linear attention, router gates, shared expert gates, shared expert projections, and MTP paths. Header-derived stored bytes are therefore authoritative for the bound: routed expert tensors total 43.487723520 GB and divide exactly into 0.084936960 GB per expert index across 512 expert indexes; fixed ordinary text traffic totals 4.107630080 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing routed expert weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6051842087737511, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed I32 INT4 routed expert weights, BF16 scales, I64 shape metadata, and unquantized BF16 tensors. AWQ dequantization, activation traffic, router compute, expert compute, recurrent-state writes, and framework scheduling are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32 for targeted Linear tensors. Because attention, linear-attention, router, shared-expert, embedding, and head tensors are ignored by the recipe, the resident package is not an ideal flat 0.5 bytes per parameter. The profile keeps full-attention KV cache as BF16 because the config has no KV-cache quantization scheme." }, "evidence": [ { "label": "Cyankiwi Qwen3-Coder-Next AWQ 4-bit API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-Next-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "downloads", "base_model_proof", "license", "pipeline", "serving", "commit_sha" ], "notes": "At commit fa29333988617831d44a4bb68ee3ebde2616346c, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_next, conversational, endpoints_compatible, compressed-tensors, region:us, and base_model:Qwen/Qwen3-Coder-Next tags. Current downloads are 118028. The API config reports Qwen3NextForCausalLM/qwen3_next and compressed-tensors quantization metadata." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/raw/fa29333988617831d44a4bb68ee3ebde2616346c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 4-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/raw/fa29333988617831d44a4bb68ee3ebde2616346c/recipe.yaml", "source_type": "config", "supports": [ "quantization_scope", "serving", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe applies AWQ to Linear targets with 4-bit symmetric integer weights, group_size 32, group strategy, and MSE observer. It ignores model.embed_tokens, linear_attn tensors, norms, shared_expert tensors, shared_expert_gate, mlp.gate, router tensors, self_attn tensors, lm_head, and MTP paths. That leaves routed mlp.experts tensors as the quantized expert traffic." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no differences in the audited architecture fields between the cyankiwi AWQ config and the official BF16 base config. The cyankiwi repo adds compressed-tensors quantization_config and keeps dtype bfloat16 while preserving the base model geometry." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/resolve/fa29333988617831d44a4bb68ee3ebde2616346c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The resolved safetensors index metadata reports total_size 48217683456 bytes. Direct range-read safetensors headers across all ten shards found 221847 tensors totaling the same 48.217683456 GB: I32 packed weights 38.654705664 GB, BF16 tensors/scales 9.561798144 GB, and I64 shape tensors 0.001179648 GB. The resident-only model.embed_tokens.weight tensor is 0.622329856 GB, leaving 47.595353600 GB main resident bytes. Routed expert tensors total 43.487723520 GB across 221184 tensors and divide exactly to 0.084936960 GB per expert index. Fixed ordinary text traffic, including linear attention, full attention, routers, shared expert, shared expert gates, norms, and lm_head, totals 4.107630080 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned config, compressed-tensors AWQ recipe, official BF16 base config comparison, resolved safetensors index, direct shard-header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the generated ideal 4-bit estimate, which missed that most always-on Qwen3-Next language tensors remain BF16, missed the shared expert, omitted fixed DeltaNet state, and understated both resident footprint and ordinary decode traffic." }, { "id": "cyankiwi--qwen3-coder-next-awq-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-Coder-Next-AWQ-8bit", "title": "Cyankiwi Qwen3-Coder-Next AWQ 8-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 8-bit package of Qwen3-Coder-Next.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, pinned config comparison, compressed-tensors recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The API metadata records Qwen/Qwen3-Coder-Next as the quantized base model. Manual comparison found matching memory-relevant architecture fields between the cyankiwi config and the audited official BF16 base config: architecture, model type, layer count, full_attention_interval, hidden size, expert geometry, attention heads, KV heads, head dimensions, linear-attention state geometry, max positions, vocabulary size, and untied embeddings. The cyankiwi config adds compressed-tensors AWQ quantization metadata while retaining dtype bfloat16." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 86.87238912, "main_resident_weight_gb": 86.250059264, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.160434432, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "resolved safetensors index total_size and direct shard-header stored bytes for the compressed-tensors AWQ package", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed BF16 language tensors plus expected distinct routed expert tensor groups", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. The compressed-tensors recipe excludes shared expert, shared expert gate, and router gate tensors from quantization, so they remain BF16 and are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "The AWQ package only quantizes routed expert Linear tensors. The recipe excludes embeddings, lm_head, norms, full attention, linear attention, router gates, shared expert gates, shared expert projections, and MTP paths. Header-derived stored bytes are therefore authoritative for the bound: routed expert tensors total 82.142429184 GB and divide exactly into 0.160434432 GB per expert index across 512 expert indexes; fixed ordinary text traffic totals 4.107630080 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing routed expert weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1.090342677326, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int8-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed I32 INT8 routed expert weights, BF16 scales, I64 shape metadata, and unquantized BF16 tensors. AWQ dequantization, activation traffic, router compute, expert compute, recurrent-state writes, and framework scheduling are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized 8-bit integer weights with group_size 32 for targeted Linear tensors. Because attention, linear-attention, router, shared-expert, embedding, and head tensors are ignored by the recipe, the resident package is larger than the logical 80B model-card count would suggest under a flat 1 byte per parameter assumption. The profile keeps full-attention KV cache as BF16 because the config has no KV-cache quantization scheme." }, "evidence": [ { "label": "Cyankiwi Qwen3-Coder-Next AWQ 8-bit API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-Next-AWQ-8bit", "source_type": "model_card", "supports": [ "repo", "downloads", "base_model_proof", "license", "pipeline", "serving", "commit_sha" ], "notes": "At commit a5a598b8531a5ef032db7242cf7f5b49ab2eada6, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_next, conversational, endpoints_compatible, compressed-tensors, region:us, and base_model:Qwen/Qwen3-Coder-Next tags. Current downloads are 108320. The API config reports Qwen3NextForCausalLM/qwen3_next and compressed-tensors quantization metadata." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 8-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/raw/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and compressed-tensors pack-quantized 8-bit integer weights with group_size 32." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 8-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/raw/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/recipe.yaml", "source_type": "config", "supports": [ "quantization_scope", "serving", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe applies AWQ to Linear targets with 8-bit symmetric integer weights, group_size 32, group strategy, and MSE observer. It ignores model.embed_tokens, linear_attn tensors, norms, shared_expert tensors, shared_expert_gate, mlp.gate, router tensors, self_attn tensors, lm_head, and MTP paths. That leaves routed mlp.experts tensors as the quantized expert traffic." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no differences in the audited architecture fields between the cyankiwi AWQ config and the official BF16 base config after excluding layer_types, dtype field naming, and compressed-tensors quantization_config. The target stores layer_types explicitly while the base derives the same pattern from full_attention_interval 4." }, { "label": "Cyankiwi Qwen3-Coder-Next AWQ 8-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/resolve/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The resolved safetensors index metadata reports total_size 86872389120 bytes and total_parameters 24108399360 storage-accounting elements. Direct range-read safetensors headers across all 18 shards found 221847 tensors totaling the same 86.872389120 GB: I32 packed weights 77.309411328 GB, BF16 tensors/scales 9.561798144 GB, and I64 shape tensors 0.001179648 GB. The resident-only model.embed_tokens.weight tensor is 0.622329856 GB, leaving 86.250059264 GB main resident bytes. Routed expert tensors total 82.142429184 GB across 221184 tensors and divide exactly to 0.160434432 GB per expert index. Fixed ordinary text traffic, including linear attention, full attention, routers, shared expert, shared expert gates, norms, and lm_head, totals 4.107630080 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned config, compressed-tensors AWQ recipe, official BF16 base config comparison, resolved safetensors index, direct shard-header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the generated ideal 8-bit estimate, which missed that many always-on Qwen3-Next language tensors remain BF16, missed the shared expert, omitted fixed DeltaNet state, and understated both resident footprint and ordinary decode traffic." }, { "id": "cyankiwi--qwen3-next-80b-a3b-instruct-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit", "title": "cyankiwi Qwen3-Next 80B A3B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3-Next 80B A3B Instruct.", "model_family": "qwen3-next-moe-instruct-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-Next-80B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, audited BF16 base profile, and direct base-config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3-Next-80B-A3B-Instruct as its quantized base model. Manual comparison found no differences in the checked model geometry fields between the audited BF16 base config and this artifact after excluding quantization metadata, dtype packaging, and Transformers bookkeeping fields." }, "architecture": { "canonical_architecture_id": "qwen3-next-80b-a3b-instruct", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 49.203395072, "main_resident_weight_gb": 47.5953536, "auxiliary_resident_weight_gb": 1.608041472, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.08493696, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding and top-level MTP tensors", "auxiliary_scope": "model.embed_tokens.weight and top-level mtp tensors are resident for the package but not swept for ordinary non-speculative text decode", "shared_expert_notes": "The config records shared_expert_intermediate_size 512 and the model card states one shared expert. The compressed-tensors recipe and served ignore list leave shared_expert and shared_expert_gate modules uncompressed; shared expert traffic is included in fixed_weight_gb rather than routed expert traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 512 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors. Top-level MTP tensors are charged as resident package bytes but ordinary non-speculative decode does not sweep them." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary non-speculative text decode; MTP/speculative decoding requires a separate workload/profile path." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5872139018215551, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-next-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, MTP speculative execution, recurrent-state writes, and prefill scheduling are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 runtime dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by API safetensors storage-accounting elements; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3-Next 80B A3B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA b213ae173e2ec9eebe9191b0ee2686c3173d520c, the API records a public Apache-2.0 text-generation repo derived from Qwen/Qwen3-Next-80B-A3B-Instruct, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 79985. The API safetensors block reports I64: 150528, I32: 78920024064, BF16: 4871089408, total: 83791264000 storage-accounting elements." }, { "label": "cyankiwi Qwen3-Next 80B A3B Instruct AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit/raw/b213ae173e2ec9eebe9191b0ee2686c3173d520c/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, BF16 runtime dtype, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, 48 layers, full_attention_interval 4, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 512 experts, 10 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, one MTP layer, and kv_cache_scheme null." }, { "label": "Qwen3-Next 80B A3B Instruct BF16 profile and config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/raw/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in audited architecture fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config, dtype packaging, and Transformers bookkeeping fields. The existing BF16 profile records the same Qwen3-Next hybrid KV/state adapter and resident-only MTP convention used here." }, { "label": "cyankiwi Qwen3-Next 80B A3B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit/resolve/b213ae173e2ec9eebe9191b0ee2686c3173d520c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 10 indexed shards. Stored tensor payloads sum to the index total_size, 49.203395072 GB, across 226472 tensors: BF16 9.742178816 GB, I32 39.460012032 GB, and I64 0.001204224 GB. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 47.595353600 GB. Auxiliary resident tensors, defined as top-level mtp tensors plus model.embed_tokens.weight, sum to 1.608041472 GB. Routed expert tensors sum to 43.487723520 GB and divide exactly into 512 uniform expert indexes of 0.084936960 GB. Fixed ordinary text traffic sums to 4.107630080 GB." }, { "label": "cyankiwi Qwen3-Next 80B A3B Instruct AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit/raw/b213ae173e2ec9eebe9191b0ee2686c3173d520c/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention modules, norms, shared experts, shared_expert_gate, mlp.gate, router modules, self-attention modules, MTP modules, and lm_head. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, BF16 base profile/config comparison, quantization recipe, direct safetensors index resolution, range-read safetensors shard headers, and the existing Transformers qwen3_next runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared-expert count, and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, MTP, linear-attention, attention, shared-expert, gate, router, and norm modules." }, { "id": "cyankiwi--qwen3-next-80b-a3b-thinking-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit", "title": "cyankiwi Qwen3-Next 80B A3B Thinking AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the cyankiwi compressed-tensors AWQ 4-bit package of Qwen3-Next 80B A3B Thinking.", "model_family": "qwen3-next-moe-thinking-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-Next-80B-A3B-Thinking", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, current BF16 base config, and direct base-config comparison", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3-Next-80B-A3B-Thinking as its quantized base model. Manual comparison found matching audited geometry fields between the current BF16 base config and this artifact after excluding quantization metadata, dtype packaging, generated layer_types bookkeeping, and Transformers metadata fields." }, "architecture": { "canonical_architecture_id": "qwen3-next-80b-a3b-thinking", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 49.203395072, "main_resident_weight_gb": 47.5953536, "auxiliary_resident_weight_gb": 1.608041472, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.08493696, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding and top-level MTP tensors", "auxiliary_scope": "model.embed_tokens.weight and top-level mtp tensors are resident for the package but not swept for ordinary non-speculative text decode", "shared_expert_notes": "The config records shared_expert_intermediate_size 512 and the model card states one shared expert. The compressed-tensors recipe and served ignore list leave shared_expert and shared_expert_gate modules uncompressed; shared expert traffic is included in fixed_weight_gb rather than routed expert traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 512 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors. Top-level MTP tensors are charged as resident package bytes but ordinary non-speculative decode does not sweep them." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary non-speculative text decode; MTP/speculative decoding requires a separate workload/profile path." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5872139018215551, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-next-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, MTP speculative execution, recurrent-state writes, and prefill scheduling are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 runtime dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by API safetensors storage-accounting elements; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3-Next 80B A3B Thinking AWQ model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 1c7d46f509dc053c42befba435aecb39ce6a7921, the API records a public Apache-2.0 text-generation repo derived from Qwen/Qwen3-Next-80B-A3B-Thinking, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 105276. The API safetensors block reports I64: 150528, I32: 78920024064, BF16: 4871089408, total: 83791264000 storage-accounting elements." }, { "label": "cyankiwi Qwen3-Next 80B A3B Thinking AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit/raw/1c7d46f509dc053c42befba435aecb39ce6a7921/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, BF16 runtime dtype, compressed-tensors pack-quantized 4-bit integer weights with group_size 32 and symmetric quantization, 48 layers, full_attention_interval 4, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 512 experts, 10 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, one MTP layer, and kv_cache_scheme null." }, { "label": "Qwen3-Next 80B A3B Thinking BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking/raw/e502dd4100cc68c0de57643fd4317ec93a128670/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in audited architecture fields between the current BF16 base config and this compressed-tensors artifact after excluding quantization_config, dtype field naming, generated layer_types, and Transformers bookkeeping fields. The base API reports BF16 safetensors total 81324862720 parameters." }, { "label": "cyankiwi Qwen3-Next 80B A3B Thinking AWQ safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit/resolve/1c7d46f509dc053c42befba435aecb39ce6a7921/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 10 indexed shards. Stored tensor payloads sum to the index total_size, 49.203395072 GB, across 226472 tensors: BF16 9.742178816 GB, I32 39.460012032 GB, and I64 0.001204224 GB. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 47.595353600 GB. Auxiliary resident tensors, defined as top-level mtp tensors plus model.embed_tokens.weight, sum to 1.608041472 GB. Routed expert tensors sum to 43.487723520 GB and divide exactly into 512 uniform expert indexes of 0.084936960 GB. Fixed ordinary text traffic sums to 4.107630080 GB." }, { "label": "cyankiwi Qwen3-Next 80B A3B Thinking AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit/raw/1c7d46f509dc053c42befba435aecb39ce6a7921/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, linear-attention modules, norms, shared experts, shared_expert_gate, mlp.gate, router modules, self-attention modules, MTP modules, and lm_head. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, current BF16 base config comparison, quantization recipe, direct safetensors index resolution, range-read safetensors shard headers, and the existing Transformers qwen3_next runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared-expert count, modeled Qwen3-Next as full-context attention only, and undercounted compressed-tensors side tensors plus unquantized embeddings, output head, MTP, linear-attention, attention, shared-expert, gate, router, and norm modules." }, { "id": "cyankiwi--qwen3-vl-2b-instruct-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit", "title": "cyankiwi Qwen3 VL 2B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for cyankiwi's compressed-tensors AWQ 4-bit package of Qwen3-VL 2B Instruct.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-2B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, recipe.yaml, target config, and Qwen base config comparison", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a compressed-tensors AWQ derivative of Qwen/Qwen3-VL-2B-Instruct. Manual comparison found matching Qwen3VLForConditionalGeneration shape, text layer geometry, visual tower geometry, tied embeddings, BF16 text dtype, and 262144 max-position embeddings. The AWQ target config omits the base config's rope_scaling and rope_theta metadata, so this profile uses the target repo config directly for bounds fields and relies on the pinned Qwen config only as architecture evidence." }, "architecture": { "canonical_architecture_id": "qwen3-vl-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.171572616, "swept_params_b": 1.76461556, "auxiliary_resident_params_b": 0.406957056, "resident_weight_gb": 2.229218368, "swept_weight_gb": 1.415304256, "auxiliary_resident_weight_gb": 0.813914112, "resident_parameter_scope": "hf_api_logical_compressed_tensors_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "model.language_model safetensors headers, including packed AWQ tensors, BF16 scales, I64 shape side tensors, BF16 norms, and tied embed_tokens/output projection", "auxiliary_scope": "model.visual tensors are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "HF API logical compressed-tensors parameters are used for parameter counts, with I32 packed tensors counted as unpacked 4-bit logical weights. Exact range-read safetensors header bytes drive memory footprint and per-token traffic. The config records tie_word_embeddings true and the header has no lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The target text config records full-context attention geometry with 28 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.0265456248505207, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored compressed-tensors bytes: packed I32 int4 payloads, BF16 scale tensors, ignored BF16 modules, and I64 shape side tensors from safetensors headers. AWQ dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "recipe.yaml and config.json record AWQ 4-bit weights with group size 32, symmetric group quantization, no activation quantization, no KV cache quantization scheme, ignored visual modules, ignored embeddings/norms in the recipe, ignored lm_head, and BF16 text dtype. This profile charges BF16 KV/state scalars from the served text config." }, "evidence": [ { "label": "cyankiwi Qwen3 VL 2B AWQ 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At repo SHA 66408a5298047012e0059208c6306a3a64e597af, the live HF API response records an Apache-2.0 public non-gated image-text-to-text repo with base_model Qwen/Qwen3-VL-2B-Instruct, compressed-tensors metadata, region:us, 351144 downloads, and safetensors logical parameters I64: 392, I32: 1409286144, BF16: 762286080, total: 2171572616." }, { "label": "cyankiwi Qwen3 VL 2B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/raw/66408a5298047012e0059208c6306a3a64e597af/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The target config records Qwen3VLForConditionalGeneration, compressed-tensors pack-quantized AWQ weights, group_size 32, symmetric group strategy, no KV cache scheme, tie_word_embeddings true, BF16 top-level and text dtype, 28 text layers, hidden_size 2048, intermediate_size 6144, 16 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, vocab_size 151936, and a resident 24-layer visual tower with hidden_size 1024." }, { "label": "Qwen3 VL 2B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/89644892e4d85e24eaac8bacfd4f463576704203/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching checked architecture fields except that the target AWQ config adds top-level dtype bfloat16 and omits the base config's rope_scaling and rope_theta metadata. Shape, context, text dtype, attention geometry, vocabulary size, tied embeddings, and visual tower geometry match." }, { "label": "cyankiwi Qwen3 VL 2B AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/raw/66408a5298047012e0059208c6306a3a64e597af/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope", "serving" ], "notes": "recipe.yaml records AWQModifier targets [Linear], 4-bit int weights, symmetric group quantization, group_size 32, no activation quantization, MSE observer, and ignore patterns for embed_tokens, norms, model.visual.*, and lm_head." }, { "label": "cyankiwi Qwen3 VL 2B AWQ safetensors header", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/resolve/66408a5298047012e0059208c6306a3a64e597af/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "Range-reading model.safetensors found a 125968-byte header and 1017 tensors totaling 2.229218368 GB, exactly matching the linked object size of 2229344344 bytes after adding the 8-byte length prefix and header. Stored bytes split into I32 0.704643072 GB, BF16 1.524572160 GB, and I64 0.000003136 GB. Logical compressed-tensors parameters total 2.171572616B after counting I32 packed int4 payloads as unpacked logical weights. model.language_model tensors total 1.764615560B logical parameters / 1.415304256 GB and are swept for ordinary text decode, including the tied embedding/output projection. model.visual tensors total 0.406957056B logical parameters / 0.813914112 GB resident-only for ordinary text decode. The header has model.language_model.embed_tokens.weight and no lm_head.weight." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, HF CLI repo info/file listing, the model card, pinned served compressed-tensors config, pinned Qwen base config comparison, recipe.yaml, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, including the generated ideal 1.0858 GB resident estimate and flat generated active-weight estimate." }, { "id": "cyankiwi--qwen3-vl-30b-a3b-instruct-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit", "title": "cyankiwi Qwen3 VL 30B A3B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for cyankiwi's compressed-tensors AWQ 4-bit package of Qwen3-VL 30B A3B Instruct.", "model_family": "qwen3-vl-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, BF16 base config comparison, quantization recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The compressed-tensors artifact records Qwen/Qwen3-VL-30B-A3B-Instruct as its quantized base model. Manual comparison found matching checked top-level, text_config, and vision_config geometry fields between the public BF16 base config and this artifact. The artifact adds compressed-tensors AWQ metadata and leaves the Qwen3VLMoeForConditionalGeneration geometry unchanged." }, "architecture": { "canonical_architecture_id": "qwen3-vl-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 19.164870112, "main_resident_weight_gb": 17.46527744, "auxiliary_resident_weight_gb": 1.699592672, "fixed_weight_gb": 1.157528576, "routed_expert_weight_gb": 0.127404288, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through model.language_model excluding embed_tokens plus lm_head, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept as full matrices for each ordinary text decode token", "shared_expert_notes": "The text config records no shared expert. Router/gate tensors are uncompressed BF16 and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 128 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 48 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state cache, or KV quantization scheme." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.61681380800292, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-qwen3-vl-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, and scheduler overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, BF16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by the audited BF16 base 31.070754032B model parameter identity; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "cyankiwi Qwen3 VL 30B A3B Instruct AWQ 4-bit API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3-VL-30B-A3B-Instruct, with compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 149879. The API safetensors block reports I64: 37248, I32: 29896998912, BF16: 2108036336, and total: 5845198448 storage-accounting elements; expanding the packed I32 tensors gives 32.005072496B logical storage elements, but exact header bytes drive this profile." }, { "label": "cyankiwi Qwen3 VL 30B A3B Instruct AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit/raw/3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The served config records Qwen3VLMoeForConditionalGeneration, qwen3_vl_moe_text, tie_word_embeddings false, 48 text layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, decoder_sparse_step 1, no shared expert, norm_topk_prob true, max_position_embeddings 262144, vocab_size 151936, a 27-layer resident visual tower, and compressed-tensors pack-quantized 4-bit integer weights with group_size 32, symmetric quantization, MSE observer, and kv_cache_scheme null." }, { "label": "Qwen3 VL 30B A3B Instruct base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in the checked top-level, text_config, and vision_config geometry fields between the BF16 base config and this compressed-tensors artifact after excluding quantization_config and packaging metadata." }, { "label": "cyankiwi Qwen3 VL 30B A3B Instruct AWQ 4-bit safetensors index and shard headers", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit/resolve/3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four indexed shards. Stored tensor payloads sum to the index total_size, 19.164870112 GB, across 56466 tensors: BF16 4.216072672 GB, I32 14.948499456 GB, and I64 0.000297984 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 17.465277440 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1.699592672 GB. Routed expert tensors sum to 16.307748864 GB and divide exactly into 128 uniform expert indexes of 0.127404288 GB. Fixed ordinary text traffic sums to 1.157528576 GB." }, { "label": "cyankiwi Qwen3 VL 30B A3B Instruct AWQ 4-bit recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit/raw/3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records AWQ 4-bit integer weights with group_size 32 and MSE observer, targeting Linear modules while ignoring embeddings, input/post-attention layer norms, final norms, model.visual.*, router gate modules, and lm_head. The served config and safetensors headers remain authoritative for exact stored tensor bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, pinned compressed-tensors config, BF16 base config comparison, quantization recipe, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as an ideal flat 4-bit MoE package and undercounted compressed-tensors side tensors plus unquantized visual, embedding, output, router, norm, and scale tensors." }, { "id": "cyankiwi--qwen3-vl-4b-instruct-awq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit", "title": "cyankiwi Qwen3 VL 4B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for cyankiwi's compressed-tensors AWQ 4-bit package of Qwen3-VL 4B Instruct.", "model_family": "qwen3-vl-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, recipe.yaml, target config, audited BF16 base config comparison, and safetensors header review", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a compressed-tensors AWQ derivative of Qwen/Qwen3-VL-4B-Instruct. Manual comparison found matching Qwen3VLForConditionalGeneration shape, text layer geometry, visual tower geometry, BF16 text dtype, M-RoPE metadata, vocabulary size, and 262144 max-position embeddings. The target repo changes root tie_word_embeddings to false, stores a separate BF16 lm_head.weight, and adds compressed-tensors quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.940313592, "swept_params_b": 4.13600972, "auxiliary_resident_params_b": 0.804303872, "resident_weight_gb": 4.430656448, "swept_weight_gb": 2.822048704, "auxiliary_resident_weight_gb": 1.608607744, "resident_parameter_scope": "hf_api_logical_compressed_tensors_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "ordinary text decode excludes model.language_model.embed_tokens.weight input lookup and includes language layer tensors, final norm, and lm_head.weight output projection", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept as full matrices for each ordinary generated text token", "notes": "HF API logical compressed-tensors parameters are used for parameter counts, with I32 packed tensors counted as unpacked 4-bit logical weights. Exact range-read safetensors header bytes drive memory footprint and per-token traffic. The root config records tie_word_embeddings false and the header stores both model.language_model.embed_tokens.weight and lm_head.weight as BF16 [151936, 2560] tensors. Therefore the input embedding is resident-only for ordinary decode, while lm_head.weight remains in swept output-projection traffic. Visual tensors remain resident-only for this text-decode profile." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The target text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.8968370864502805, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-compressed-tensors-awq-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored compressed-tensors bytes: packed I32 int4 payloads, BF16 scale tensors, ignored BF16 modules, BF16 visual/input/head tensors, and I64 shape side tensors from safetensors headers. AWQ dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "recipe.yaml and config.json record AWQ 4-bit weights with group_size 32, symmetric group quantization, no activation quantization, no KV cache quantization scheme, ignored visual modules, ignored embeddings/norms in the recipe, ignored lm_head, and BF16 text dtype. This profile charges BF16 KV/state scalars from the served text config." }, "evidence": [ { "label": "cyankiwi Qwen3 VL 4B AWQ 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At repo SHA 5ab8b8a020ffd84a86008d4f628d796f1827f182, the live HF API response records an Apache-2.0 public non-gated image-text-to-text repo with base_model Qwen/Qwen3-VL-4B-Instruct, compressed-tensors metadata, region:us, 116951 downloads, and safetensors logical parameters I64: 504, I32: 3633315840, BF16: 1306997248, total: 4940313592." }, { "label": "cyankiwi Qwen3 VL 4B AWQ 4-bit config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit/raw/5ab8b8a020ffd84a86008d4f628d796f1827f182/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The target config records Qwen3VLForConditionalGeneration, root tie_word_embeddings false, text_config tie_word_embeddings true, compressed-tensors pack-quantized AWQ weights, group_size 32, symmetric group strategy, no KV cache scheme, BF16 top-level and text dtype, 36 text layers, hidden_size 2560, intermediate_size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, vocab_size 151936, and a resident 24-layer visual tower with hidden_size 1024." }, { "label": "Qwen3 VL 4B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching checked architecture fields: Qwen3VLForConditionalGeneration, qwen3_vl_text model type, BF16 text dtype, 36 text layers, hidden size, intermediate size, attention heads, KV heads, head dimension, 262144 max context, vocabulary size, M-RoPE metadata, and visual tower geometry. The base BF16 config has root tie_word_embeddings true; this AWQ artifact stores an untied lm_head.weight." }, { "label": "cyankiwi Qwen3 VL 4B AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit/raw/5ab8b8a020ffd84a86008d4f628d796f1827f182/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope", "serving" ], "notes": "recipe.yaml records AWQModifier targets [Linear], 4-bit int weights, symmetric group quantization, group_size 32, no activation quantization, MSE observer, duo_scaling true, and ignore patterns for embed_tokens, input_layernorm, mlp.gate, post_attention_layernorm, norm, model.visual.*, and lm_head." }, { "label": "cyankiwi Qwen3 VL 4B AWQ safetensors header", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit/resolve/5ab8b8a020ffd84a86008d4f628d796f1827f182/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Range-reading model.safetensors found a 152168-byte header and 1218 tensors totaling 4.430656448 GB, exactly matching the linked object size of 4430808624 bytes after adding the 8-byte length prefix and header. Stored bytes split into I32 1.816657920 GB, BF16 2.613994496 GB, and I64 0.000004032 GB. Logical compressed-tensors parameters total 4.940313592B after counting I32 packed int4 payloads as unpacked logical weights. model.visual tensors total 0.415347712B logical parameters / 0.830695424 GB resident-only. model.language_model.embed_tokens.weight and lm_head.weight are both BF16 [151936, 2560] tensors contributing 0.388956160B logical parameters / 0.777912320 GB each. Ordinary text swept traffic is language layers plus final norm plus lm_head.weight, totaling 4.136009720B logical parameters / 2.822048704 GB. The resident-only ordinary text subset is model.visual plus input embedding, totaling 0.804303872B logical parameters / 1.608607744 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI repo info/file listing, model card, pinned served compressed-tensors config, pinned Qwen base config comparison, recipe.yaml, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, including the generated ideal 2.4701 GB resident estimate. It is an ordinary cached text-decode memory-side profile for the AWQ artifact." }, { "id": "davidau--qwen3-6-27b-heretic-uncensored-finetune-neo-code-di-imatrix-max-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF", "title": "DavidAU Qwen3.6 27B Heretic NEO CODE GGUF Q8_0", "summary": "Audited memory-side text-decode bounds profile for the HF API-selected Q8_0 GGUF artifact of DavidAU's Qwen3.6 27B Heretic NEO CODE finetune.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, full-precision base config, Qwen3.6 parent config comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo records the DavidAU full-precision repo as its quantized base. That base repo records trohrbaugh/Qwen3.6-27B-heretic-ara as its finetune base, while its served config matches the audited Qwen/Qwen3.6-27B geometry on all checked memory-relevant fields. The selected GGUF header records the same 64-layer qwen35 geometry and 262144-token context." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b-heretic-neo-code", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 29.86634176, "swept_weight_gb": 28.504487936, "auxiliary_resident_weight_gb": 1.361853824, "resident_parameter_scope": "selected Q8_0 GGUF linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q8_0 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for each ordinary text decode token; separate mmproj sidecar GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.6-27B-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. The selected linked file is 29.866341760 GB. Header tensor spans total 29.855348736 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010993024 GB. The selected main GGUF contains output.weight, token_embd.weight, output_norm.weight, and ordinary blk.0-63 tensors, with no MTP, mmproj, vision, or draft tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The base config and GGUF metadata record 64 ordinary text layers with every fourth layer using full attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected main Q8_0 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 1.1104381121963467, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q8-0-high-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is a Q8_0 GGUF with BF16 output components. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "DavidAU Qwen3.6 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit af8c40cc94da8d4675a78e5578c293df3ac68226 records a public Apache-2.0 image-text-to-text GGUF repo with base_model DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking, 235458 downloads, region:us, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 29866341760. The API totalFileSize exactly matches the HIGH-Q8_0 linked artifact, so this profile targets that artifact." }, { "label": "DavidAU Qwen3.6 27B GGUF model card", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking, dual-imatrix GGUF quantization, separate mmproj files for images, 256K context, a special Q8_0 quant with BF16 components, and detailed benchmark material for Q6_K and Q8_0." }, { "label": "DavidAU Qwen3.6 27B full-precision base config", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking/raw/2bca4e1caef0db358b478988446b95993226965d/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The full-precision base config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and a resident vision config." }, { "label": "Qwen3.6 27B parent config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison against the audited Qwen/Qwen3.6-27B parent config found matching checked memory-relevant fields: architecture, model type, language_model_only flag, tied-embedding setting, BF16 text dtype, 64 layers, hidden size, intermediate size, attention geometry, linear-attention geometry, Mamba SSM dtype, vocabulary size, and 262144 max positions." }, { "label": "DavidAU Qwen3.6 27B GGUF linked-object checks", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF/tree/af8c40cc94da8d4675a78e5578c293df3ac68226", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HF CLI linked-object metadata records Q8_0 HIGH 29.866341760 GB, Q6_K 22.396159360 GB, Q5_K_M 19.542148480 GB, Q5_K_S 18.990663040 GB, Q4_K_M 16.861398400 GB, Q4_K_S 15.900312960 GB, IQ4_NL 16.115926400 GB, IQ4_XS 15.397242240 GB, IQ3_M 12.891186560 GB, IQ2_M 10.486343040 GB, and separate mmproj sidecars of 0.927607360 GB to 1.842940480 GB. Q8_0 HIGH exactly matches API gguf.totalFileSize." }, { "label": "DavidAU Qwen3.6 27B Q8_0 GGUF range-read tensor index", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF/resolve/af8c40cc94da8d4675a78e5578c293df3ac68226/Qwen3.6-27B-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 42 metadata entries and 851 tensors. The linked file is 29.866341760 GB. Tensor spans sum to 29.855348736 GB: output.weight plus output_norm.weight 2.542817280 GB, token_embd.weight 1.350860800 GB, and ordinary blk.0-63 tensors 25.961670656 GB. Metadata/tokenizer/header/file overhead accounts for 0.010993024 GB. Tensor spans split into Q8_0 27.134197760 GB, BF16 2.710568960 GB, and F32 0.010582016 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, and no MTP, mmproj, or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, full-precision base config, Qwen3.6 parent config comparison, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected Q8_0 artifact." }, "notes": "Use this profile for the API-selected Q8_0 GGUF artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or Q4_K_M serving behavior unless the workload profile explicitly selects those separate files." }, { "id": "davidau--qwen3-6-40b-claude-4-6-opus-deckard-heretic-uncensored-thinking-neo-code-di-imatrix-max-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF", "title": "DavidAU Qwen3.6 40B Deck Opus NEO CODE GGUF Q8_0", "summary": "Audited memory-side text-decode bounds profile for the HF API-selected Q8_0 GGUF artifact of DavidAU's 96-layer Qwen3.6-derived model.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, full-precision base config, Qwen3.6 parent config comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo records the DavidAU full-precision repo as its quantized base. That base repo records Qwen/Qwen3.6-27B as its base but changes the text stack from 64 to 96 layers while keeping the Qwen3.6/Qwen3.5 hybrid attention and DeltaNet geometry. The selected GGUF header records the same 96-layer qwen35 geometry and 262144-token context." }, "architecture": { "canonical_architecture_id": "qwen3-6-40b-deck-opus-neo-code", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 39.072596736, "swept_params_b": 37.801198336, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 42.847202336, "swept_weight_gb": 41.485323264, "auxiliary_resident_weight_gb": 1.361879072, "resident_parameter_scope": "selected Q8_0 GGUF linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.95 tensors from the selected Q8_0 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for each ordinary text decode token; separate mmproj sidecar GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.6-40B-Deck-Opus-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. The selected linked file is 42.847202336 GB. Header tensor spans total 42.836184064 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.011018272 GB. The selected main GGUF contains output.weight, token_embd.weight, output_norm.weight, and ordinary blk.0-95 tensors, with no MTP, mmproj, vision, or draft tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 24, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The base config and GGUF metadata record 96 ordinary text layers with every fourth layer using full attention, giving 24 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.232390656, "read_gb_per_output_token": 0.232390656, "state_formula": "72 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected main Q8_0 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 1.0966049332606098, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q8-0-high-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is a Q8_0 GGUF with BF16 output and value tensors. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "DavidAU Qwen3.6 40B GGUF HF API metadata", "url": "https://huggingface.co/api/models/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 8f6654b98bf23b4d18374b7f72312cfba61d66db records a public Apache-2.0 image-text-to-text GGUF repo with base_model DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking, 504420 downloads, region:us, GGUF architecture qwen35, 262144 context length, gguf.total 39072596736, and gguf.totalFileSize 42847202336. The API totalFileSize exactly matches the HIGH-Q8_0 linked artifact, so this profile targets that artifact." }, { "label": "DavidAU Qwen3.6 40B GGUF model card", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking, dual-imatrix GGUF quantization, separate mmproj files for images, and local-app examples that also expose a Q4_K_M alias. The llama-cpp-python example explicitly names Qwen3.6-40B-Deck-Opus-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf." }, { "label": "DavidAU Qwen3.6 40B full-precision base config", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking/raw/bc23dff65597927a7f43e74b7a0deb6e49d773da/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The full-precision base config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, BF16 text config, 96 text layers, full_attention_interval 4, 24 full-attention layers, 72 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and a resident vision config." }, { "label": "Qwen3.6 27B parent config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison against the Qwen/Qwen3.6-27B parent found matching Qwen3.6 hybrid geometry fields except for num_hidden_layers: the parent records 64 layers while the DavidAU base records 96 layers." }, { "label": "DavidAU Qwen3.6 40B GGUF linked-object HEAD checks", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF/tree/8f6654b98bf23b4d18374b7f72312cfba61d66db", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Q8_0 HIGH 42.847202336 GB, Q6_K 32.391855136 GB, Q5_K_M 28.195267616 GB, Q5_K_S 27.368039456 GB, Q4_K_M 24.253604896 GB, Q4_K_S 22.801818656 GB, IQ4_NL 23.135396896 GB, IQ4_XS 22.077236256 GB, IQ3_M 18.382715936 GB, IQ2_M 14.775450656 GB, and separate mmproj sidecars of 0.927607360 GB to 1.842940480 GB. Q8_0 HIGH exactly matches API gguf.totalFileSize." }, { "label": "DavidAU Qwen3.6 40B Q8_0 GGUF range-read tensor index", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF/resolve/8f6654b98bf23b4d18374b7f72312cfba61d66db/Qwen3.6-40B-Deck-Opus-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 43 metadata entries and 1275 tensors. The linked file is 42.847202336 GB. Tensor spans sum to 42.836184064 GB: output.weight 2.5427968 GB, token_embd.weight 1.3508608 GB, blk.* tensors 38.942505984 GB, and output_norm.weight 0.00002048 GB. Metadata/tokenizer/header/file overhead accounts for 0.011018272 GB. Tensor spans split into Q8_0 40.02586624 GB, BF16 2.79445504 GB, and F32 0.015862784 GB. The header records qwen35.block_count 96, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no MTP, mmproj, or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, full-precision base config, Qwen3.6 parent config comparison, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected Q8_0 artifact." }, "notes": "Use this profile for the API-selected Q8_0 GGUF artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or Q4_K_M serving behavior unless the workload profile explicitly selects those separate files." }, { "id": "decart-ai--kimi-k2-7-code-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "decart-ai/Kimi-K2.7-Code-NVFP4", "title": "Decart Kimi K2.7 Code NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for Decart's ModelOpt NVFP4 Kimi K2.7 Code serving artifact.", "model_family": "kimi-k2-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-K2.7-Code", "relation": "quantized", "source": "Hugging Face model card base_model metadata, Decart model card, served config, hf_quant_config, base-config comparison, custom-code hash comparison, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The repo metadata and card identify moonshotai/Kimi-K2.7-Code as the base model. Checked architecture fields match the audited base config after excluding quantization metadata and the derivative text_config.model_type label. configuration_deepseek.py, configuration_kimi_k25.py, modeling_deepseek.py, and modeling_kimi_k25.py are byte-identical to the audited base repo files." }, "architecture": { "canonical_architecture_id": "kimi-k2-7-code", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 595.148146656, "main_resident_weight_gb": 591.857048576, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 21.095607296, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60 and lm_head, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. ModelOpt quantizes routed expert Linear targets but excludes self-attention, shared_experts, lm_head, vision_tower, and mm_projector; shared expert tensors are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the artifact mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and small F32 tensors. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but constructs expanded key_states and value_states before past_key_values.update. Decart's hf_quant_config records FP8 KV cache, so Bounds Engine v1 charges expanded FP8 K/V cache streams for this serving artifact." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image/video encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8_expanded_key_value_cache", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_expanded_key_value_cache", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed NVFP4 payloads, F8 scale tensors, BF16/F32 fixed tensors, and FP8 expanded K/V cache bytes. Dequantization, activation traffic, vision/video prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records ModelOpt NVFP4 weights/activations with group_size 16 and the hf_quant_config records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "Decart Kimi K2.7 Code NVFP4 API metadata", "url": "https://huggingface.co/api/models/decart-ai/Kimi-K2.7-Code-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving" ], "notes": "At commit b16b4af9d9255c114bf3d5f124dc1ba12cbf812c, the API reports a public text-generation repo with license other / modified MIT metadata, Model Optimizer library, ModelOpt/NVFP4 tags, custom_code, base_model moonshotai/Kimi-K2.7-Code, region:us, current downloads 352463, and a safetensors index with total_parameters 519536364528 and total_size 595148146656 bytes." }, { "label": "Decart Kimi K2.7 Code NVFP4 model card", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "model_family", "max_context_tokens" ], "notes": "The card identifies the repo as a Decart community reproduction of a ModelOpt NVFP4 quantization of Moonshot AI Kimi K2.7 Code. It states routed-expert linear layers are NVFP4 with FP8 KV cache, while attention/MLA, shared experts, the layer-0 dense MLP, lm_head, vision_tower, and mm_projector remain BF16. It also states the precision split matches nvidia/Kimi-K2.6-NVFP4 and documents vLLM serving on NVIDIA Blackwell." }, { "label": "Decart Kimi K2.7 Code NVFP4 config", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/raw/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, ModelOpt NVFP4 group_size 16, quantized Linear targets, and exclusions for layer 0, self-attention, shared experts, lm_head, vision_tower, and mm_projector." }, { "label": "Decart Kimi K2.7 Code NVFP4 quantization config", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/raw/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "fixed_weight_gb" ], "notes": "hf_quant_config records producer modelopt 0.37.0-decart-k2.7, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes language_model.lm_head, language_model.model.layers.0*, all language_model.model.layers.*.self_attn modules, all shared_experts modules, mm_projector, and vision_tower from weight quantization." }, { "label": "Decart Kimi K2.7 Code NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/resolve/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 595148146656 bytes across 60 shards and 277670 tensors. Direct range-read shard headers match the index total and sum exactly to 595.148146656 GB: U8 507.343011840 GB, F8_E4M3 63.417876480 GB, BF16 24.386705376 GB, and F32 0.000552960 GB. Language tensors excluding input embeddings total 591.857048576 GB; vision_tower contributes 0.833732064 GB, mm_projector 0.108555776 GB, and input embedding 2.348810240 GB, so resident-only auxiliary tensors sum to 3.291098080 GB. Routed expert tensors total 570.761441280 GB, exactly 1.486357920 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095607296 GB." }, { "label": "Decart Kimi K2.7 Code NVFP4 custom-code hash comparison", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/tree/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual SHA-256 comparison found Decart's configuration_deepseek.py, configuration_kimi_k25.py, modeling_deepseek.py, and modeling_kimi_k25.py match the audited moonshotai/Kimi-K2.7-Code files exactly. The audited runtime review for the base found DeepseekV3Attention expands K/V into key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_values.update." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the Decart model card, served config, hf_quant_config, base-config comparison, custom-code hash comparison, range-read safetensors shard headers, and the existing audited Kimi K2.7 custom-code runtime review." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the Decart ModelOpt NVFP4 artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "deepreinforce-ai--ornith-1-0-35b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepreinforce-ai/Ornith-1.0-35B-GGUF", "title": "DeepReinforce Ornith 1.0 35B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of Ornith 1.0 35B.", "model_family": "ornith-qwen3.5-moe-gguf", "base_model_proof": { "base_model": "deepreinforce-ai/Ornith-1.0-35B", "relation": "derived_package", "source": "GGUF model card, live HF API GGUF metadata, non-GGUF Ornith config/API metadata, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo card documents Ornith-1.0-35B, and the non-GGUF Ornith config records Qwen3_5MoeForConditionalGeneration with the same memory-relevant geometry as the selected GGUF header: 40 text blocks, full_attention_interval 4, 10 full-attention layers, 30 DeltaNet linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 256 routed experts, 8 experts per token, one shared expert, and 262144 context length. The card says the 35B MoE is post-trained on top of Qwen 3.5." }, "architecture": { "canonical_architecture_id": "ornith-1-0-35b-qwen35moe", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 69.3766368, "main_resident_weight_gb": 68.348529152, "auxiliary_resident_weight_gb": 1.028107648, "fixed_weight_gb": 3.924019712, "routed_expert_weight_gb": 0.25165824, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected ornith-1.0-35b-bf16.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected BF16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for each ordinary text decode token", "shared_expert_notes": "The GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B marketing parameters. The selected BF16 GGUF stores BF16 tensors plus tiny F32 tensors. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The non-GGUF config and selected GGUF metadata record 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Ornith/Qwen3.5 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 main GGUF artifact. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and speculative paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-moe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected BF16 GGUF artifact. GGUF loader overhead, kernels, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The API-selected artifact is ornith-1.0-35b-bf16.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param records nominal BF16 tensor payload size." }, "evidence": [ { "label": "Ornith 1.0 35B GGUF HF API metadata", "url": "https://huggingface.co/api/models/deepreinforce-ai/Ornith-1.0-35B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit c2e1703039380de4ce6820e97afd185682d3c16c records a public MIT-licensed text-generation GGUF repo with 436780 downloads, region:us, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 69376636800. The API totalFileSize matches ornith-1.0-35b-bf16.gguf, so this profile targets that artifact." }, { "label": "Ornith 1.0 35B GGUF model card", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF/raw/c2e1703039380de4ce6820e97afd185682d3c16c/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "runtime_format" ], "notes": "The card documents Ornith-1.0-35B as a 35B MoE reasoning model for agentic coding, post-trained on top of Qwen 3.5, MIT licensed, and intended for recent vLLM, SGLang, Transformers, or GGUF runtimes. It lists the 35B model as part of a family with 9B/31B dense and 35B/397B MoE variants." }, { "label": "Ornith 1.0 35B non-GGUF API metadata", "url": "https://huggingface.co/api/models/deepreinforce-ai/Ornith-1.0-35B", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family" ], "notes": "The non-GGUF repo is public, MIT-licensed, text-generation, and records qwen3_5_moe plus image-text-to-text tags at commit 5df2ed3f675c7beaa490328cc70bb573b65fb660. Its API safetensors block reports BF16 parameters, while the pinned config gives the memory-relevant architecture used for compatibility checks." }, { "label": "Ornith 1.0 35B non-GGUF config", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/raw/5df2ed3f675c7beaa490328cc70bb573b65fb660/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, and 262144 max position embeddings." }, { "label": "Ornith 1.0 35B BF16 GGUF linked object and range-read tensor index", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF/resolve/c2e1703039380de4ce6820e97afd185682d3c16c/ornith-1.0-35b-bf16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 40 metadata entries and 733 tensors. The linked file is 69.376636800 GB. Tensor spans sum to 69.365647872 GB; metadata/tokenizer/header/file overhead accounts for 0.010988928 GB. Tensor spans split into BF16 69.276794880 GB and F32 0.088852992 GB. token_embd.weight is 1.017118720 GB and resident-only; output.weight is a separate 1.017118720 GB swept tensor. Routed expert tensors sum to 64.424509440 GB, or 0.251658240 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 3.924019712 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found Q4_K_M 21.166757760 GB, Q5_K_M 24.729130848 GB, Q6_K 28.514152288 GB, Q8_0 36.903138880 GB, and BF16 69.376636800 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, non-GGUF config/API metadata, selected linked file sizes, a direct GGUF header/tensor-index range read of the API-selected BF16 artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected Ornith 1.0 35B BF16 main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "deepseek-ai--deepseek-coder-6-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/deepseek-coder-6.7b-instruct", "title": "DeepSeek Coder 6.7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 DeepSeek Coder 6.7B Instruct repo.", "model_family": "deepseek-coder-llama-dense", "base_model_proof": { "base_model": "deepseek-ai/deepseek-coder-6.7b-base", "relation": "finetune", "source": "DeepSeek Coder model card and direct config comparison", "config_compatible": true, "notes": "The model card says deepseek-coder-6.7b-instruct is initialized from deepseek-coder-6.7b-base. Current served configs match on LlamaForCausalLM architecture, hidden/intermediate sizes, layer count, attention/KV heads, context length, vocabulary size, BF16 dtype, untied embeddings, linear RoPE scaling, and RoPE theta." }, "architecture": { "canonical_architecture_id": "deepseek-coder-6.7b", "max_context_tokens": 16384, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.740512768, "swept_params_b": 6.608392192, "auxiliary_resident_params_b": 0.132120576, "resident_weight_gb": 13.481025536, "swept_weight_gb": 13.216784384, "auxiliary_resident_weight_gb": 0.264241152, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers plus model.norm plus separate lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens.weight is resident for the package but not swept as a full matrix for ordinary text decode", "notes": "The config records tie_word_embeddings false, and the safetensors headers contain a separate lm_head.weight. Ordinary cached text decode sweeps the decoder layers, final norm, and the separate output projection. The input embedding table remains resident but is not charged as per-token swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 Llama decoder layers, 32 attention heads, 32 key/value heads, hidden size 4096, no sliding-window setting, and 16384 max position embeddings, so this profile charges full-context BF16 K and V streams for cached text decode." }, "notes": "DeepSeek Coder 6.7B Instruct uses LlamaForCausalLM with linear RoPE scaling for a 16K context window. This profile models ordinary cached text decode only; prefill, tokenizer overhead, and code-generation quality are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-llama-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges exact BF16 stored tensors and BF16 KV streams from the served config and tensor headers.", "notes": "The config records torch_dtype bfloat16, and direct safetensors headers record only BF16 tensor payloads." }, "evidence": [ { "label": "DeepSeek Coder 6.7B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-coder-6.7b-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "serving", "total_params_b" ], "notes": "At repo SHA e5d64addd26a6a1db0f9b863abf6ee3141936807, the API records a public non-gated text-generation repo with transformers, pytorch, safetensors, llama, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 382398. The API safetensors block reports BF16: 6740512768. The model card says DeepSeek Coder 6.7B Instruct is initialized from DeepSeek Coder 6.7B Base and fine-tuned on instruction data, with a 16K window and commercial use under the DeepSeek Coder model license." }, { "label": "DeepSeek Coder 6.7B Instruct config", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/raw/e5d64addd26a6a1db0f9b863abf6ee3141936807/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, BF16 dtype, hidden size 4096, intermediate size 11008, 32 hidden layers, 32 attention heads, 32 KV heads, 16384 max position embeddings, linear RoPE scaling factor 4, rope_theta 100000, vocab size 32256, and tie_word_embeddings false." }, { "label": "DeepSeek Coder 6.7B Base config comparison", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base/raw/ce2207a8bfef3ee92bd7dd4cc31c52cfa0046912/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "config_compatible" ], "notes": "A direct current config comparison against the base repo found matching architecture fields, including layer count, hidden/intermediate sizes, attention/KV heads, context length, vocabulary size, BF16 dtype, untied embeddings, linear RoPE scaling, and RoPE theta." }, { "label": "DeepSeek Coder 6.7B Instruct safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/resolve/e5d64addd26a6a1db0f9b863abf6ee3141936807/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split" ], "notes": "The index records total_size 13481025536 bytes across two shards. Direct range-read safetensors headers match the index exactly and found 291 BF16 tensors totaling 6740512768 params / 13.481025536 GB. model.layers tensors total 12.952535040 GB, model.norm.weight is 0.000008192 GB, lm_head.weight is a separate 0.264241152 GB tensor, and model.embed_tokens.weight is a separate 0.264241152 GB tensor. Ordinary text swept tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 6608392192 params / 13.216784384 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served config, direct base-config comparison, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the generated dense estimate by replacing rounded weight fields with exact header bytes and by separating resident input embeddings from the swept output projection." }, { "id": "deepseek-ai--deepseek-coder-7b-instruct-v1-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/deepseek-coder-7b-instruct-v1.5", "title": "DeepSeek Coder 7B Instruct v1.5 BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 DeepSeek Coder 7B Instruct v1.5 repo.", "model_family": "deepseek-coder-llama-dense", "architecture": { "canonical_architecture_id": "deepseek-coder-7b-v1.5", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.910365696, "swept_params_b": 6.490935296, "auxiliary_resident_params_b": 0.4194304, "resident_weight_gb": 13.820731392, "swept_weight_gb": 12.981870592, "auxiliary_resident_weight_gb": 0.8388608, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers plus model.norm plus separate lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens.weight is resident for the package but not swept as a full matrix for ordinary text decode", "notes": "The config records tie_word_embeddings false, and the safetensors headers contain a separate lm_head.weight. Ordinary text decode sweeps the decoder layers, final norm, and the separate output projection. The input embedding table remains resident but is not charged as per-token swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 30, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 30 Llama decoder layers, 32 attention heads, 32 key/value heads, hidden size 4096, no sliding-window setting, and 4096 max positions, so this profile charges full-context BF16 K and V streams for cached text decode." }, "notes": "DeepSeek Coder 7B Instruct v1.5 uses LlamaForCausalLM. This profile models ordinary cached text decode only; prefill, tokenizer overhead, and code-generation quality are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-llama-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges exact BF16 stored tensors and BF16 KV streams from the served config and tensor headers.", "notes": "The config records torch_dtype bfloat16, and direct safetensors headers record only BF16 tensor payloads." }, "evidence": [ { "label": "DeepSeek Coder 7B Instruct v1.5 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-coder-7b-instruct-v1.5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "serving", "total_params_b" ], "notes": "At repo SHA 2a050a4c59d687a85324d32e147517992117ed30, the API records a public text-generation repo with transformers, safetensors, llama, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 549396. The API safetensors block reports BF16: 6910365696. The model card describes DeepSeek Coder 7B Instruct v1.5 as continuing from DeepSeek LLM 7B with a 4K window, then instruction fine-tuned, and says commercial use is supported under the DeepSeek Coder model license." }, { "label": "DeepSeek Coder 7B Instruct v1.5 config", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5/raw/2a050a4c59d687a85324d32e147517992117ed30/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, BF16 dtype, hidden size 4096, intermediate size 11008, 30 hidden layers, 32 attention heads, 32 KV heads, 4096 max position embeddings, rope_theta 10000, vocab size 102400, and tie_word_embeddings false." }, { "label": "DeepSeek Coder 7B Instruct v1.5 safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5/resolve/2a050a4c59d687a85324d32e147517992117ed30/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split" ], "notes": "The index records total_size 13820731392 bytes across three shards. Direct range-read safetensors headers match the index exactly and found 273 BF16 tensors totaling 6910365696 params / 13.820731392 GB. model.layers tensors total 12.143001600 GB, model.norm.weight is 0.000008192 GB, lm_head.weight is a separate 0.838860800 GB tensor, and model.embed_tokens.weight is a separate 0.838860800 GB tensor. Ordinary text swept tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 6490935296 params / 12.981870592 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served config, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the generated dense estimate by replacing rounded weight fields with exact header bytes and by separating resident input embeddings from the swept output projection." }, { "id": "deepseek-ai--deepseek-coder-v2-lite-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "title": "DeepSeek Coder V2 Lite Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 DeepSeek-Coder-V2-Lite-Instruct repo.", "model_family": "deepseek-coder-v2-lite-moe", "architecture": { "canonical_architecture_id": "deepseek-coder-v2-lite", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.412968448, "main_resident_weight_gb": 30.993538048, "auxiliary_resident_weight_gb": 0.4194304, "fixed_weight_gb": 2.203835392, "routed_expert_weight_gb": 0.449839104, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary Transformers text decode through layers 0-26, model.norm.weight, and lm_head.weight, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 2. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded 16B/2.4B model-card parameters. Routed expert tensors are byte-uniform across 64 expert indexes; routed_expert_weight_gb is the total routed expert byte count divided by 64." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers attention path constructs cached key_states with 16 heads and q_head_dim 192, composed of 128 no-RoPE dimensions plus 64 RoPE dimensions." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers eager attention path stores value_states with 16 heads and v_head_dim 128 before past_key_value.update." } ], "notes": "The architecture is MLA-style internally, but the bundled custom Transformers code expands compressed_kv into key_states and value_states before cache update. Bounds Engine v1 therefore charges expanded BF16 K/V cache streams for all 27 decoder layers. This profile does not assume runtime-specific latent MLA cache compression." }, "notes": "The served config records 27 ordinary hidden layers, with layer 0 dense and layers 1-26 MoE. There is no separate MTP/next-token-prediction layer in this checkpoint." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, state writes, and kernel efficiency remain outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and the model card's Transformers example loads with torch_dtype=torch.bfloat16. The vLLM example requires an external PR and is not treated as evidence for a runtime-specific compressed-cache implementation in this profile." }, "evidence": [ { "label": "DeepSeek-Coder-V2-Lite-Instruct model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "At repo SHA e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11, the API records a public/non-gated text-generation repo with deepseek_v2, custom_code, safetensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 1094936. The API safetensors block reports BF16: 15706484224 and total: 15706484224. The card describes DeepSeek-Coder-V2-Lite-Instruct as a 16B total / 2.4B activated MoE model with 128K context and gives Transformers BF16 and vLLM examples." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records DeepseekV2ForCausalLM, model_type deepseek_v2, BF16 dtype, 27 hidden layers, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10944, moe_intermediate_size 1408, 16 attention heads, kv_lora_rank 512, q_lora_rank null, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, tie_word_embeddings false, YaRN rope scaling factor 40, and 163840 max position embeddings." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct custom modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV2DecoderLayer builds layer 0 as dense MLP and layers 1-26 as DeepseekV2MoE. DeepseekV2MoE routes to top-k experts and adds shared_experts every MoE layer. DeepseekV2Attention computes compressed_kv internally but expands it into key_states with 16 heads x 192 dims and value_states with 16 heads x 128 dims before past_key_value.update in the eager path." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/resolve/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four shards. Stored tensor bytes sum to 31.412968448 GB, all BF16, matching the index metadata total_size and API parameter count. model.embed_tokens.weight is 0.419430400 GB resident-only. Ordinary main tensors excluding input embeddings sum to 30.993538048 GB. Routed expert tensors in model.layers.1-26.mlp.experts.* sum to 28.789702656 GB, exactly 0.449839104 GB per expert index. Fixed ordinary-decode traffic including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head is 2.203835392 GB." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct license section", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "source_type": "model_card", "supports": [ "license" ], "notes": "The card metadata records license other, license_name deepseek-license, and license_link LICENSE. The license section states the model is subject to the DeepSeek model license and that the DeepSeek-Coder-V2 series supports commercial use." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, pinned custom modeling code, safetensors index, direct range-read safetensors shard headers, and the license section." }, "notes": "This profile replaces the generated metadata estimate, which undercounted active MoE traffic and used a rounded full-KV coefficient. It is an ordinary text-decode profile for the official BF16 Transformers artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "deepseek-ai--deepseek-r1-0528-qwen3-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "title": "DeepSeek R1 0528 Qwen3 8B BF16", "summary": "Audited memory-side bounds profile for the official BF16 DeepSeek R1 0528 Qwen3 8B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B-Base", "relation": "finetune", "source": "DeepSeek model card plus direct base-config comparison", "config_compatible": false, "notes": "The model card says DeepSeek-R1-0528 was used to post-train Qwen3 8B Base. Manual comparison with Qwen/Qwen3-8B-Base config found matching core tensor geometry, but DeepSeek serves 131072 max position embeddings while the base config serves 32768, so this profile uses the DeepSeek config directly." }, "architecture": { "canonical_architecture_id": "deepseek-r1-0528-qwen3-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 16.38147072, "swept_weight_gb": 15.136811008, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 399 BF16 tensors totaling 8190735360 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served DeepSeek repo config rather than deriving structure from the model name or from upstream Qwen defaults." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 0528 Qwen3 8B model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The card identifies the model as DeepSeek-R1-0528 post-training Qwen3 8B Base and states the model is MIT licensed." }, { "label": "DeepSeek R1 0528 Qwen3 8B API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format", "pipeline" ], "notes": "At commit 6e8885a6ff5c1dc5201574c8fd700323f23c25fa, the API safetensors block records BF16: 8190735360 and total: 8190735360, which this profile stores as 8.19073536B resident parameters." }, { "label": "DeepSeek R1 0528 Qwen3 8B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 36 layers, 8 KV heads, 128 head dimension, 131072 max position embeddings, no sliding window, and untied embeddings." }, { "label": "DeepSeek R1 0528 Qwen3 8B safetensors headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index lists two safetensors shards with total_size 16381470720 bytes. Range-read shard headers record 399 BF16 tensors totaling 8190735360 parameters and 16.38147072 GB. model.embed_tokens.weight has shape [151936, 4096] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 15.136811008 GB." }, { "label": "Qwen3 8B Base config", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/raw/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, hidden size, layer count, attention geometry, dtype, vocabulary size, and untied embeddings; max_position_embeddings differs, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from pinned DeepSeek config, pinned model card, current HF API metadata, pinned safetensors index, direct safetensors header range reads, and direct Qwen3-8B-Base config comparison." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "deepseek-ai--deepseek-r1-0528", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-0528", "title": "DeepSeek R1 0528 FP8", "summary": "Audited memory-side bounds profile for the official FP8 DeepSeek R1 0528 repo.", "model_family": "deepseek-v3-moe", "architecture": { "canonical_architecture_id": "deepseek-r1-0528", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 688.57483936, "main_resident_weight_gb": 673.150611808, "auxiliary_resident_weight_gb": 15.424227552, "fixed_weight_gb": 19.082195296, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Expected-distinct routing is applied to the 256 uniform routed experts across the main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "MLA cache coefficient from config: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent/rope cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "DeepSeek R1 0528 uses MLA rather than full key/value heads. Bounds Engine v1 represents the latent cache as compressed_state with coefficients derived from the config dimensions." }, "notes": "The official 0528 checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary decode traffic excludes that auxiliary layer but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 quantization and BF16 model dtype; safetensors headers record F8_E4M3, BF16, and F32 tensors." }, "evidence": [ { "label": "DeepSeek R1 0528 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-0528", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 4236a6af538feda4548eca9ab308586007567f52, the API reports an MIT text-generation repo with FP8 tag and safetensors parameters split across BF16: 3918786560, F8_E4M3: 680571043840, and F32: 41555600 tensors." }, { "label": "DeepSeek R1 0528 config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/4236a6af538feda4548eca9ab308586007567f52/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 256 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 quantization, kv_lora_rank 512, qk_rope_head_dim 64, and 163840 max position embeddings." }, { "label": "DeepSeek R1 0528 safetensors index", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/4236a6af538feda4548eca9ab308586007567f52/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensors sum to 688.57483936 GB: F8_E4M3 680.57104384 GB, BF16 7.83757312 GB, and F32 0.1662224 GB. Main tensors in layers 0-60 sum to 673.150611808 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic sums to 19.082195296 GB. The auxiliary next-token prediction layer 61 sums to 15.424227552 GB resident-only." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from config, HF API metadata, safetensors index, and range-read safetensors header parameter/byte splits." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without metadata guessing." }, { "id": "deepseek-ai--deepseek-r1-distill-llama-70b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "title": "DeepSeek R1 Distill Llama 70B BF16", "summary": "Audited memory-side bounds profile for the official BF16 DeepSeek R1 Distill Llama 70B repo.", "model_family": "llama3.3-dense", "architecture": { "canonical_architecture_id": "deepseek-r1-distill-llama-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 141.107412992, "swept_weight_gb": 139.006066688, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 70553706496 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 80 layers, 8 KV heads, 128 head dimension, use_cache true, and no sliding-window setting. The audited Transformers Llama attention path updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "Dense LlamaForCausalLM profile using the served DeepSeek repo config rather than deriving structure from the gated Meta Llama source model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Llama 70B API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit b1c0b44b4369b597ad119a196caf79a9c40e141e, the API records a public MIT text-generation repo with transformers, safetensors, llama, conversational, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 571006. The API safetensors block records BF16: 70553706496 and total: 70553706496." }, { "label": "DeepSeek R1 Distill model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/b1c0b44b4369b597ad119a196caf79a9c40e141e/README.md", "source_type": "model_card", "supports": [ "base_model", "license", "serving" ], "notes": "The card states that DeepSeek-R1-Distill-Llama-70B is one of six dense models distilled from DeepSeek-R1, lists its base model as Llama-3.3-70B-Instruct, says the distill models were fine-tuned from open-source models using samples generated by DeepSeek-R1, notes that DeepSeek-R1-Distill models can be used like Qwen or Llama models, and states MIT licensing plus commercial use." }, { "label": "DeepSeek R1 Distill Llama 70B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/b1c0b44b4369b597ad119a196caf79a9c40e141e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records LlamaForCausalLM, bfloat16, hidden_size 8192, intermediate_size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, Llama 3 RoPE scaling factor 8 from original 8192 positions, rope_theta 500000, tie_word_embeddings false, rms_norm_eps 1e-5, vocab size 128256, and use_cache true." }, { "label": "DeepSeek R1 Distill Llama 70B safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/b1c0b44b4369b597ad119a196caf79a9c40e141e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 141107412992 bytes across 17 shards. Range-read shard headers found 723 BF16 tensors totaling 70553706496 parameters / 141.107412992 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [128256, 8192] and contributes 1050673152 parameters / 2.101346304 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 69503033344 parameters / 139.006066688 GB." }, { "label": "Transformers Llama implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/llama/modeling_llama.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found LlamaModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 80 layers for this config. LlamaAttention projects key_states and value_states, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx), supporting expanded full-context BF16 K/V cache charges for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned DeepSeek model card, pinned served config, generation config, safetensors index, direct safetensors shard header range reads, and the upstream Transformers Llama runtime implementation." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses the served DeepSeek config and tensor headers directly rather than inheriting from the gated Meta Llama source model." }, { "id": "deepseek-ai--deepseek-r1-distill-llama-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "title": "DeepSeek R1 Distill Llama 8B BF16", "summary": "Audited memory-side bounds profile for the BF16 DeepSeek R1 Distill Llama 8B repo.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B", "relation": "finetune", "source": "DeepSeek R1 model card and the served DeepSeek config", "config_compatible": false, "notes": "The model card identifies DeepSeek-R1-Distill-Llama-8B as derived from Llama3.1-8B-Base and says the distill models were fine-tuned from open-source models using samples generated by DeepSeek-R1. The card also says DeepSeek slightly changed distill configs and tokenizers, while the Meta base config is gated in this audit environment. This profile therefore uses the served DeepSeek config directly instead of inheriting unaudited base-model fields." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-llama-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 16.060522496, "swept_weight_gb": 15.009849344, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 291 BF16 tensors totaling 8030261248 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 32 layers, 8 KV heads, 128 head dimension, use_cache true, 131072 max position embeddings, and no sliding-window setting. The audited Transformers Llama attention path updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "Dense LlamaForCausalLM profile using the served DeepSeek distill config rather than deriving structure from the gated Meta Llama source model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Llama 8B API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "pipeline", "total_params_b", "weight_format", "commit_sha" ], "notes": "At commit 6a6f4aa4197940add57724a7707d069478df56b1, the API records a public non-gated MIT text-generation Transformers repo with safetensors, llama, conversational, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 305879. The API safetensors block records BF16: 8030261248 and total: 8030261248." }, { "label": "DeepSeek R1 Distill model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "serving" ], "notes": "The card says DeepSeek-R1-Distill-Llama-8B is one of six dense models distilled from DeepSeek-R1, lists its base model as Llama-3.1-8B, says the distilled models are fine-tuned based on open-source models using samples generated by DeepSeek-R1, says DeepSeek slightly changed the configs and tokenizers, and notes that DeepSeek-R1-Distill models can be used like Qwen or Llama models. It also states MIT licensing and commercial use, with the Llama 3.1 source license applying to the original base." }, { "label": "DeepSeek R1 Distill Llama 8B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records LlamaForCausalLM, bfloat16, hidden_size 4096, intermediate_size 14336, 32 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, Llama 3 RoPE scaling factor 8 from original 8192 positions, rope_theta 500000, tie_word_embeddings false, rms_norm_eps 1e-5, vocab size 128256, and use_cache true." }, { "label": "DeepSeek R1 Distill Llama 8B generation config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/generation_config.json", "source_type": "config", "supports": [ "serving" ], "notes": "The generation config records sampling defaults temperature 0.6 and top_p 0.95 with the same BOS and EOS token IDs as the served config. It does not alter memory-side decode geometry." }, { "label": "DeepSeek R1 Distill Llama 8B safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 16060522496 bytes across two shards. Range-read shard headers found 291 BF16 tensors totaling 8030261248 parameters / 16.060522496 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [128256, 4096] and contributes 525336576 parameters / 1.050673152 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7504924672 parameters / 15.009849344 GB. Linked-object HEAD checks resolved both shards to 16060556354 total bytes, leaving 33858 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Llama implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/llama/modeling_llama.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found LlamaModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 32 layers for this config. LlamaAttention projects key_states and value_states, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx), supporting expanded full-context BF16 K/V cache charges for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned DeepSeek model card, pinned served config, generation config, safetensors index, direct safetensors shard header range reads, linked-object HEAD checks, and the upstream Transformers Llama runtime implementation." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses the served DeepSeek config and tensor headers directly rather than inheriting from the gated Meta Llama source model." }, { "id": "deepseek-ai--deepseek-r1-distill-qwen-1-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "title": "DeepSeek R1 Distill Qwen 1.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 DeepSeek R1 Distill Qwen 1.5B repo.", "model_family": "qwen2-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Math-1.5B", "relation": "finetune", "source": "DeepSeek R1 model card and base config comparison", "config_compatible": false, "notes": "The model card identifies the 1.5B distill checkpoint as fine-tuned from Qwen/Qwen2.5-Math-1.5B with DeepSeek-R1 samples. The served config preserves core Qwen2.5 1.5B tensor geometry but changes max_position_embeddings to 131072 and records tie_word_embeddings false, while the base config has 4096 context and tied embeddings. This profile uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-qwen-1-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.777088, "swept_params_b": 1.543714304, "auxiliary_resident_params_b": 0.233373696, "resident_weight_gb": 3.554176, "swept_weight_gb": 3.087428608, "auxiliary_resident_weight_gb": 0.466747392, "resident_parameter_scope": "model.safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers plus model.norm plus lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The single safetensors file stores 339 BF16 tensors totaling 1777088000 parameters and 3.554176 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; ordinary text decode excludes only the input embedding lookup while keeping lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 28 layers, 2 KV heads, 128 head dimension, and use_sliding_window false, so Bounds Engine v1 charges full-context K and V streams for all layers." }, "notes": "Dense Qwen2ForCausalLM profile using the served DeepSeek distill config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Qwen 1.5B model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "serving" ], "notes": "At commit ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562, the README says DeepSeek-R1-Distill-Qwen-1.5B is based on Qwen2.5-Math-1.5B, that distill models can be used like Qwen or Llama models, and that the model weights are MIT licensed." }, { "label": "DeepSeek R1 Distill Qwen 1.5B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/raw/ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, hidden size 1536, intermediate size 8960, 28 layers, 12 attention heads, 2 KV heads, 131072 max position embeddings, sliding_window 4096, use_sliding_window false, untied embeddings, and vocab size 151936." }, { "label": "DeepSeek R1 Distill Qwen 1.5B Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "The current API response records commit ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562, 630704 downloads when audited, public non-gated transformers text-generation metadata, region:us, and safetensors parameters BF16: 1777088000." }, { "label": "DeepSeek R1 Distill Qwen 1.5B safetensors header", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/resolve/ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "A direct range read of the 38613-byte safetensors header found 339 BF16 tensors. model.layers total 2.620678144 GB, model.norm.weight totals 0.000003072 GB, lm_head.weight is a separate [151936, 1536] BF16 tensor of 0.466747392 GB, and model.embed_tokens.weight is another [151936, 1536] BF16 tensor of 0.466747392 GB." }, { "label": "Qwen2.5 Math 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching hidden size, intermediate size, layer count, attention head count, KV head count, dtype, vocab size, and model type, but different context and tied-embedding settings." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served config, model card/base-model statement, base config comparison, and a direct safetensors header range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the input embedding as swept decode traffic and used rounded parameter counts." }, { "id": "deepseek-ai--deepseek-r1-distill-qwen-14b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "title": "DeepSeek R1 Distill Qwen 14B BF16", "summary": "Audited memory-side bounds profile for the official BF16 DeepSeek R1 Distill Qwen 14B repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-14B", "relation": "finetune", "source": "DeepSeek model card plus direct base-config comparison", "config_compatible": true, "notes": "The model card identifies DeepSeek-R1-Distill-Qwen-14B as a distilled checkpoint based on Qwen2.5-14B and says the distill models were fine-tuned from open-source models using samples generated by DeepSeek-R1. Manual comparison found the served DeepSeek config matches the checked Qwen/Qwen2.5-14B tensor geometry and context fields." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-qwen2-5-14b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 29.540067328, "swept_weight_gb": 27.982931968, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 14770033664 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served DeepSeek repo config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Qwen 14B API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 1df8507178afcc1bef68cd8c393f61a886323761, the API records a public MIT repo with transformers, safetensors, endpoints_compatible, and region:us tags. Current downloads are 452833. The API safetensors block records BF16: 14770033664 and total: 14770033664. The current API response does not expose a pipeline_tag for this repo." }, { "label": "DeepSeek R1 Distill model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/raw/1df8507178afcc1bef68cd8c393f61a886323761/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "serving" ], "notes": "The card states that DeepSeek-R1-Distill-Qwen-14B is one of six dense models distilled from DeepSeek-R1, lists its base model as Qwen2.5-14B, says the distilled models are fine-tuned based on open-source models using samples generated by DeepSeek-R1, and notes that DeepSeek-R1-Distill models can be used like Qwen or Llama models. It also states MIT licensing and commercial use." }, { "label": "DeepSeek R1 Distill Qwen 14B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/raw/1df8507178afcc1bef68cd8c393f61a886323761/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, hidden_size 5120, intermediate_size 13824, 48 layers, 40 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, rope_theta 1000000, rms_norm_eps 1e-5, vocab size 152064, and 131072 max position embeddings." }, { "label": "Qwen2.5 14B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B/raw/97e1e76335b7017d8f67c08a19d103c0504298c9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, dtype, hidden size, intermediate size, layer count, attention head count, KV head count, vocabulary size, sliding_window, use_sliding_window, max_position_embeddings, rope_theta, rms_norm_eps, and tied-embedding setting between the base config and the served DeepSeek config." }, { "label": "DeepSeek R1 Distill Qwen 14B safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/raw/1df8507178afcc1bef68cd8c393f61a886323761/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 29540067328 bytes across four shards. Range-read shard headers found 579 BF16 tensors totaling 14770033664 parameters / 29.540067328 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 5120] and contributes 778567680 parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 13991465984 parameters / 27.982931968 GB. Linked-object HEAD checks resolved all four shards to 29540133872 total bytes, leaving 66544 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned DeepSeek model card, pinned served config, direct Qwen2.5 14B base config comparison, safetensors index, direct safetensors shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses the served DeepSeek config rather than deriving structure from the model name." }, { "id": "deepseek-ai--deepseek-r1-distill-qwen-32b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "title": "DeepSeek R1 Distill Qwen 32B BF16", "summary": "Audited memory-side bounds profile for the official BF16 DeepSeek R1 Distill Qwen 32B repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B", "relation": "finetune", "source": "DeepSeek model card plus direct base-config comparison", "config_compatible": true, "notes": "The model card identifies DeepSeek-R1-Distill-Qwen-32B as a distilled checkpoint based on Qwen2.5-32B and says the distill models were fine-tuned from open-source models using samples generated by DeepSeek-R1. Manual comparison found the served DeepSeek config matches the checked Qwen/Qwen2.5-32B tensor geometry and context fields." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-qwen2-5-32b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 65.527752704, "swept_weight_gb": 63.970617344, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 32763876352 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served DeepSeek repo config. The config records 131072 max position embeddings, matching the Qwen2.5 32B base config rather than the Qwen2.5 32B Instruct config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Qwen 32B model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 711ad2ea6aa40cfca18895e8aca02ab92df1a746, the API records a public MIT text-generation repo with qwen2, safetensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 826902. The API safetensors block records BF16: 32763876352 and total: 32763876352." }, { "label": "DeepSeek R1 Distill Qwen 32B model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "serving" ], "notes": "The card states that DeepSeek-R1-Distill-Qwen-32B is one of six dense models distilled from DeepSeek-R1, lists its base model as Qwen2.5-32B, says the distilled models are fine-tuned based on open-source models using samples generated by DeepSeek-R1, and notes that DeepSeek-R1-Distill models can be used like Qwen or Llama models. It also states MIT licensing and commercial use." }, { "label": "DeepSeek R1 Distill Qwen 32B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, rope_theta 1000000, rms_norm_eps 1e-5, vocab size 152064, and 131072 max position embeddings." }, { "label": "Qwen2.5 32B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B/raw/1818d35814b8319459f4bd55ed1ac8709630f003/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture, dtype, hidden size, intermediate size, layer count, attention head count, KV head count, vocabulary size, sliding_window, use_sliding_window, max_position_embeddings, rope_theta, rms_norm_eps, eos_token_id, and tied-embedding setting between the base config and the served DeepSeek config." }, { "label": "DeepSeek R1 Distill Qwen 32B safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 65527752704 bytes across 8 shards. Range-read shard headers found 771 BF16 tensors totaling 32763876352 parameters / 65.527752704 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 5120] and contributes 778567680 parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 31985308672 parameters / 63.970617344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned DeepSeek model card, pinned served config, direct Qwen2.5 32B base config comparison, safetensors index, direct safetensors shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses the served DeepSeek config rather than deriving structure from the model name." }, { "id": "deepseek-ai--deepseek-r1-distill-qwen-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "title": "DeepSeek R1 Distill Qwen 7B BF16", "summary": "Audited memory-side bounds profile for the BF16 DeepSeek R1 Distill Qwen 7B repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Math-7B", "relation": "finetune", "source": "DeepSeek R1 model card and direct Qwen2.5 Math 7B base config comparison", "config_compatible": false, "notes": "The model card identifies DeepSeek-R1-Distill-Qwen-7B as a distilled checkpoint based on Qwen2.5-Math-7B and says the distill models were fine-tuned from open-source models using samples generated by DeepSeek-R1. Manual comparison found matching tensor geometry, dtype, sliding-window settings, vocabulary size, RoPE theta, and untied embedding layout, but the served DeepSeek config sets max_position_embeddings to 131072 while the current Qwen2.5-Math-7B base config sets 4096. This profile therefore uses the served DeepSeek config directly." }, "architecture": { "canonical_architecture_id": "deepseek-r1-distill-qwen2-5-7b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.231233024, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 339 BF16 tensors totaling 7615616512 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 28 layers, 4 KV heads, 128 head dimension, and use_sliding_window false, so Bounds Engine v1 charges full-context K and V streams for all layers." }, "notes": "Dense Qwen2ForCausalLM profile using the served DeepSeek distill config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "DeepSeek R1 Distill Qwen 7B API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "pipeline", "total_params_b", "weight_format", "commit_sha" ], "notes": "At commit 916b56a44061fd5cd7d6a8fb632557ed4f724f60, the API records a public non-gated MIT text-generation Transformers repo with safetensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 329897. The API safetensors block records BF16: 7615616512 and total: 7615616512." }, { "label": "DeepSeek R1 Distill model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/raw/916b56a44061fd5cd7d6a8fb632557ed4f724f60/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "serving" ], "notes": "The card says DeepSeek-R1-Distill-Qwen-7B is one of six dense models distilled from DeepSeek-R1, lists its base model as Qwen2.5-Math-7B, says the distilled models are fine-tuned based on open-source models using samples generated by DeepSeek-R1, and notes that DeepSeek-R1-Distill models can be used like Qwen or Llama models. It also states MIT licensing and commercial use." }, { "label": "DeepSeek R1 Distill Qwen 7B config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/raw/916b56a44061fd5cd7d6a8fb632557ed4f724f60/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, hidden_size 3584, intermediate_size 18944, 28 layers, 28 attention heads, 4 KV heads, sliding_window 4096, use_sliding_window false, tie_word_embeddings false, rope_theta 10000, rms_norm_eps 1e-6, vocab size 152064, use_mrope false, and 131072 max position embeddings." }, { "label": "Qwen2.5 Math 7B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B/raw/b101308fe89651ea5ce025f25317fea6fc07e96e/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, dtype, hidden size, intermediate size, layer count, attention head count, KV head count, vocabulary size, sliding_window, use_sliding_window, RoPE theta, rms_norm_eps, and tied-embedding setting between the base config and the served DeepSeek config. The served DeepSeek config changes max_position_embeddings from 4096 to 131072." }, { "label": "DeepSeek R1 Distill Qwen 7B safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/raw/916b56a44061fd5cd7d6a8fb632557ed4f724f60/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 15231233024 bytes across two shards. Range-read shard headers found 339 BF16 tensors totaling 7615616512 parameters / 15.231233024 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB. Linked-object HEAD checks resolved both shards to 15231271850 total bytes, leaving 38826 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned DeepSeek model card, pinned served config, direct Qwen2.5 Math 7B base config comparison, safetensors index, direct safetensors shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses the served DeepSeek config rather than deriving structure from the model name." }, { "id": "deepseek-ai--deepseek-r1", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-R1", "title": "DeepSeek R1 FP8", "summary": "Audited memory-side bounds profile for the official FP8 DeepSeek R1 repo.", "model_family": "deepseek-v3-moe", "architecture": { "canonical_architecture_id": "deepseek-r1", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 688.57483936, "main_resident_weight_gb": 673.150611808, "auxiliary_resident_weight_gb": 15.424227552, "fixed_weight_gb": 19.082195296, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Expected-distinct routing is applied to the 256 uniform routed experts across the main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "MLA cache coefficient from config: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent/rope cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "DeepSeek R1 uses MLA rather than full key/value heads. Bounds Engine v1 represents the latent cache as compressed_state with coefficients derived from the config dimensions." }, "notes": "The official checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary decode traffic excludes that auxiliary layer but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 quantization and BF16 model dtype; safetensors headers record F8_E4M3, BF16, and F32 tensors." }, "evidence": [ { "label": "DeepSeek R1 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "The API reports an MIT text-generation repo with FP8 tag and safetensors parameters split across BF16, F8_E4M3, and F32 tensors." }, { "label": "DeepSeek R1 config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, 256 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 quantization, kv_lora_rank 512, qk_rope_head_dim 64, and 163840 max position embeddings." }, { "label": "DeepSeek R1 safetensors index", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensors sum to 688.57483936 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, or 2.554954752 GB per expert index. Main fixed ordinary-decode traffic sums to 19.082195296 GB. The auxiliary next-token prediction layer 61 sums to 15.424227552 GB resident-only." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from config, HF API metadata, and safetensors header parameter/byte splits." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without metadata guessing." }, { "id": "deepseek-ai--deepseek-v2-lite-chat", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V2-Lite-Chat", "title": "DeepSeek V2 Lite Chat BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 DeepSeek-V2-Lite-Chat repo.", "model_family": "deepseek-v2-lite-moe", "architecture": { "canonical_architecture_id": "deepseek-v2-lite", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.412968448, "main_resident_weight_gb": 30.993538048, "auxiliary_resident_weight_gb": 0.4194304, "fixed_weight_gb": 2.203835392, "routed_expert_weight_gb": 0.449839104, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary Transformers text decode through layers 0-26, model.norm.weight, and lm_head.weight, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 2. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded 16B/2.4B model-card parameters. Routed expert tensors are byte-uniform across 64 expert indexes; routed_expert_weight_gb is the total routed expert byte count divided by 64." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers attention path constructs cached key_states with 16 heads and q_head_dim 192, composed of 128 no-RoPE dimensions plus 64 RoPE dimensions." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers eager attention path stores value_states with 16 heads and v_head_dim 128 before past_key_value.update." } ], "notes": "The architecture is MLA-style internally, but the bundled custom Transformers code expands compressed_kv into key_states and value_states before cache update. Bounds Engine v1 therefore charges expanded BF16 K/V cache streams for all 27 decoder layers. This profile does not assume runtime-specific latent MLA cache compression." }, "notes": "The served config records 27 ordinary hidden layers, with layer 0 dense and layers 1-26 MoE. There is no separate MTP/next-token-prediction layer in this checkpoint." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, state writes, and kernel efficiency remain outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and the model card's Transformers example loads with torch_dtype=torch.bfloat16. The vLLM example references DeepSeek's custom serving path and is not treated as evidence for a runtime-specific compressed-cache implementation in this profile." }, "evidence": [ { "label": "DeepSeek-V2-Lite-Chat model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V2-Lite-Chat", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "At repo SHA 85864749cd611b4353ce1decdb286193298f64c7, the API records a public/non-gated text-generation repo with deepseek_v2, custom_code, safetensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 1126300. The API safetensors block reports BF16: 15706484224 and total: 15706484224. The card describes DeepSeek-V2-Lite-Chat as a 16B total / 2.4B activated SFT chat MoE model and gives Transformers BF16 and vLLM examples." }, { "label": "DeepSeek-V2-Lite-Chat config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/raw/85864749cd611b4353ce1decdb286193298f64c7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records DeepseekV2ForCausalLM, model_type deepseek_v2, BF16 dtype, 27 hidden layers, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10944, moe_intermediate_size 1408, 16 attention heads, kv_lora_rank 512, q_lora_rank null, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, tie_word_embeddings false, YaRN rope scaling factor 40, and 163840 max position embeddings." }, { "label": "DeepSeek-V2-Lite-Chat custom modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/raw/85864749cd611b4353ce1decdb286193298f64c7/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV2DecoderLayer builds layer 0 as dense MLP and layers 1-26 as DeepseekV2MoE. DeepseekV2MoE routes to top-k experts and adds shared_experts every MoE layer. DeepseekV2Attention computes compressed_kv internally but expands it into key_states with 16 heads x 192 dims and value_states with 16 heads x 128 dims before past_key_value.update in the eager path. The pinned file is byte-identical to the already audited DeepSeek-Coder-V2-Lite-Instruct modeling file." }, { "label": "DeepSeek-V2-Lite-Chat safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/resolve/85864749cd611b4353ce1decdb286193298f64c7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four shards. Stored tensor bytes sum to 31.412968448 GB, all BF16, matching the index metadata total_size and API parameter count. model.embed_tokens.weight is 0.419430400 GB resident-only. Ordinary main tensors excluding input embeddings sum to 30.993538048 GB. Routed expert tensors in model.layers.1-26.mlp.experts.* sum to 28.789702656 GB, exactly 0.449839104 GB per expert index. Fixed ordinary-decode traffic including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head is 2.203835392 GB." }, { "label": "DeepSeek-V2-Lite-Chat license section", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat", "source_type": "model_card", "supports": [ "license" ], "notes": "The card metadata records license other, license_name deepseek, and license_link LICENSE-MODEL. The license section states the DeepSeek-V2 Base/Chat models are subject to the model license and that the DeepSeek-V2 series supports commercial use." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, pinned custom modeling code, safetensors index, direct range-read safetensors shard headers, the license section, and a direct structured comparison against the already audited DeepSeek-Coder-V2-Lite-Instruct config/modeling files." }, "notes": "This profile replaces the generated metadata estimate, which undercounted active MoE traffic and used a rounded full-KV coefficient. It is an ordinary text-decode profile for the official BF16 Transformers artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "deepseek-ai--deepseek-v2-lite", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V2-Lite", "title": "DeepSeek V2 Lite BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 DeepSeek-V2-Lite base repo.", "model_family": "deepseek-v2-lite-moe", "architecture": { "canonical_architecture_id": "deepseek-v2-lite", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.412968448, "main_resident_weight_gb": 30.993538048, "auxiliary_resident_weight_gb": 0.4194304, "fixed_weight_gb": 2.203835392, "routed_expert_weight_gb": 0.449839104, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary Transformers text decode through layers 0-26, model.norm.weight, and lm_head.weight, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 2. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded 16B/2.4B model-card parameters. Routed expert tensors are byte-uniform across 64 expert indexes; routed_expert_weight_gb is the total routed expert byte count divided by 64." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers attention path constructs cached key_states with 16 heads and q_head_dim 192, composed of 128 no-RoPE dimensions plus 64 RoPE dimensions." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "The pinned custom Transformers eager attention path stores value_states with 16 heads and v_head_dim 128 before past_key_value.update." } ], "notes": "The architecture is MLA-style internally, but the bundled custom Transformers code expands compressed_kv into key_states and value_states before cache update. Bounds Engine v1 therefore charges expanded BF16 K/V cache streams for all 27 decoder layers. This profile does not assume runtime-specific latent MLA cache compression." }, "notes": "The served config records 27 ordinary hidden layers, with layer 0 dense and layers 1-26 MoE. There is no separate MTP/next-token-prediction layer in this checkpoint." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, state writes, and kernel efficiency remain outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and the model card's Transformers example loads with torch_dtype=torch.bfloat16. The vLLM example references DeepSeek's custom serving path and is not treated as evidence for a runtime-specific compressed-cache implementation in this profile." }, "evidence": [ { "label": "DeepSeek-V2-Lite model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V2-Lite", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "At repo SHA 604d5664dddd88a0433dbae533b7fe9472482de0, the API records a public/non-gated text-generation repo with deepseek_v2, custom_code, safetensors, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 483565. The API safetensors block reports BF16: 15706484224 and total: 15706484224. The card describes DeepSeek-V2-Lite as a 16B total / 2.4B activated MoE model trained from scratch, with 27 layers, MLA, 64 routed experts, 6 activated routed experts, 2 shared experts, long-context extension, and Transformers BF16 and vLLM examples." }, { "label": "DeepSeek-V2-Lite config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/raw/604d5664dddd88a0433dbae533b7fe9472482de0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records DeepseekV2ForCausalLM, model_type deepseek_v2, BF16 dtype, 27 hidden layers, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10944, moe_intermediate_size 1408, 16 attention heads, kv_lora_rank 512, q_lora_rank null, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, tie_word_embeddings false, YaRN rope scaling factor 40, and 163840 max position embeddings. A structured comparison found these checked fields byte-equivalent to the already audited DeepSeek-V2-Lite-Chat config." }, { "label": "DeepSeek-V2-Lite custom modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/raw/604d5664dddd88a0433dbae533b7fe9472482de0/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV2DecoderLayer builds layer 0 as dense MLP and layers 1-26 as DeepseekV2MoE. DeepseekV2MoE routes to top-k experts and adds shared_experts every MoE layer. DeepseekV2Attention computes compressed_kv internally but expands it into key_states with 16 heads x 192 dims and value_states with 16 heads x 128 dims before past_key_value.update in the eager path. The pinned file is byte-identical to the already audited DeepSeek-V2-Lite-Chat and DeepSeek-Coder-V2-Lite-Instruct modeling files." }, { "label": "DeepSeek-V2-Lite safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/resolve/604d5664dddd88a0433dbae533b7fe9472482de0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all four shards. Stored tensor bytes sum to 31.412968448 GB across 5291 tensors, all BF16, matching the index metadata total_size and API parameter count. model.embed_tokens.weight is 0.419430400 GB resident-only. Ordinary main tensors excluding input embeddings sum to 30.993538048 GB. Routed expert tensors in model.layers.1-26.mlp.experts.* sum to 28.789702656 GB, exactly 0.449839104 GB per expert index. Fixed ordinary-decode traffic including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head is 2.203835392 GB. The index file is byte-identical to the already audited DeepSeek-V2-Lite-Chat index." }, { "label": "DeepSeek-V2-Lite license section", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite", "source_type": "model_card", "supports": [ "license" ], "notes": "The card metadata records license other, license_name deepseek, and a LICENSE-MODEL link. The license section states the DeepSeek-V2 Base/Chat models are subject to the model license and that the DeepSeek-V2 series supports commercial use." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, pinned custom modeling code, safetensors index, direct range-read safetensors shard headers, the license section, and byte-level comparison against the already audited DeepSeek-V2-Lite-Chat config/modeling/index files." }, "notes": "This profile replaces the generated metadata estimate, which undercounted active MoE traffic and used a rounded full-KV coefficient. It is an ordinary text-decode profile for the official BF16 Transformers artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "deepseek-ai--deepseek-v3-0324", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V3-0324", "title": "DeepSeek V3 0324 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 DeepSeek V3 0324 repo.", "model_family": "deepseek-v3-moe", "architecture": { "canonical_architecture_id": "deepseek-v3-0324", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 688.57483936, "main_resident_weight_gb": 673.150611808, "auxiliary_resident_weight_gb": 15.424227552, "fixed_weight_gb": 19.082195296, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, norm, and lm_head, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Expected-distinct routing is applied to the 256 byte-uniform routed expert indexes across the 58 main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Official DeepSeek-V3 absorb-MLA cache coefficient: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent plus RoPE cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "The official DeepSeek-V3 local inference implementation uses the absorb MLA path by default, storing kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64. The generic Hugging Face custom modeling file expands to full K/V for the Transformers cache and is not the intended local-serving path." }, "notes": "The checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary next-token-prediction layer from per-token traffic but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "official-deepseek-v3-fp8-absorb-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization, activation traffic, kernel efficiency, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records FP8 e4m3 quantization and BF16 model dtype. The official DeepSeek-V3 reference generator sets torch.set_default_dtype(torch.bfloat16), so the MLA latent/RoPE cache is charged as BF16." }, "evidence": [ { "label": "DeepSeek V3 0324 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3-0324", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit e9b33add76883f293d6bf61f6bd89b497e80e335, the API reports a public MIT text-generation repo with deepseek_v3, custom_code, fp8, and region:us tags, plus 1108710 downloads. The model card says DeepSeek-V3-0324 has exactly the same model structure as DeepSeek-V3, links the DeepSeek-V3 local-running repo, and warns that Hugging Face Transformers is not directly supported yet." }, { "label": "DeepSeek V3 0324 served config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/raw/e9b33add76883f293d6bf61f6bd89b497e80e335/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, hidden size 7168, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, FP8 e4m3 quantization, and 163840 max position embeddings." }, { "label": "DeepSeek V3 official local inference config and code", "url": "https://github.com/deepseek-ai/DeepSeek-V3/blob/9b4e9788e4a3a731f7567338ed15d3ec549ce03b/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "ordinary_decode_scope" ], "notes": "Manual review of the official DeepSeek-V3 repo at commit 9b4e9788e4a3a731f7567338ed15d3ec549ce03b found the default attn_impl absorb path. In that path MLA stores kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64, while the naive path would store expanded full K/V. The paired config_671B.json records the same 61-layer, 256-expert, 8-active-expert, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, FP8 model shape." }, { "label": "DeepSeek V3 0324 safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/resolve/e9b33add76883f293d6bf61f6bd89b497e80e335/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensor bytes sum to 688.574839360 GB: F8_E4M3 680.571043840 GB, BF16 7.837573120 GB, and F32 0.166222400 GB. Main tensors in layers 0-60 plus embeddings, norm, and lm_head sum to 673.150611808 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic, including shared experts, sums to 19.082195296 GB. The auxiliary next-token-prediction layer 61 sums to 15.424227552 GB resident-only. The index metadata total_size field records 1369.062772 GB, but direct shard headers and the API dtype parameter counts agree on 688.574839360 GB of stored bytes." }, { "label": "DeepSeek V3 0324 custom Hugging Face modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/raw/e9b33add76883f293d6bf61f6bd89b497e80e335/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "serving_caveat", "kv_adapter_boundary" ], "notes": "Manual review found that the bundled custom Transformers model expands MLA into key_states/value_states for the generic Transformers cache. The model card states Hugging Face Transformers is not directly supported yet, so this profile follows the official DeepSeek-V3 local inference path instead of the generic Transformers cache path." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the pinned served config, the model card, the official DeepSeek-V3 local inference config/code, the pinned custom HF modeling caveat, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without falling back to the stale generated full-KV estimate." }, { "id": "deepseek-ai--deepseek-v3-1", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V3.1", "title": "DeepSeek V3.1 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 DeepSeek V3.1 repo.", "model_family": "deepseek-v3-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V3.1-Base", "relation": "finetune", "source": "Hugging Face API cardData base_model field, model card text, and served base config comparison", "config_compatible": true, "notes": "The model card says DeepSeek-V3.1 is post-trained on top of DeepSeek-V3.1-Base, and the API records deepseek-ai/DeepSeek-V3.1-Base as the base model. Manual comparison found matching checked architecture, routing, MLA, context, rope, dtype, and quantization fields between the V3.1 and V3.1-Base served configs at the reviewed commits." }, "architecture": { "canonical_architecture_id": "deepseek-v3-1", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 688.57483936, "main_resident_weight_gb": 673.150611808, "auxiliary_resident_weight_gb": 15.424227552, "fixed_weight_gb": 19.082195296, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, embeddings, norm, and lm_head, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Expected-distinct routing is applied to the 256 byte-uniform routed expert indexes across the 58 main MoE layers. The model card's 671B/37B rounded counts are descriptive; bounds use exact shard bytes." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Official DeepSeek-V3 absorb-MLA cache coefficient: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent plus RoPE cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "The DeepSeek-V3.1 model card says the model structure is the same as DeepSeek-V3 and points local serving users to the DeepSeek-V3 repo. The official DeepSeek-V3 local inference implementation uses the absorb MLA path by default, storing kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64. V3.1 records UE8M0 FP8 scales for weights and activations, but the cited runtime buffers use the BF16 default dtype and no V3.2 DSA index cache is present." }, "notes": "The checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary next-token-prediction layer from per-token traffic but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "official-deepseek-v3-1-fp8-absorb-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization, UE8M0 activation scaling, activation traffic, kernel efficiency, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records FP8 e4m3 quantization, UE8M0 scale format, dynamic activation scheme, and BF16 model dtype. The official DeepSeek-V3 reference generator sets torch.set_default_dtype(torch.bfloat16), so the absorb-MLA latent/RoPE cache is charged as BF16." }, "evidence": [ { "label": "DeepSeek V3.1 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.1", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model_proof", "serving", "total_params_b" ], "notes": "At commit c0781d039fb7a1ba2abc4add0bdc293e92d2b8db, the API reports a public MIT text-generation repo with deepseek_v3, custom_code, fp8, eval-results, text-generation-inference, endpoints_compatible, region:us, and base_model deepseek-ai/DeepSeek-V3.1-Base tags, plus 250886 downloads. The API safetensors block reports BF16 3918786560, F8_E4M3 680571043840, F32 41555600, and total 684531386000 storage-accounting tensor elements." }, { "label": "DeepSeek V3.1 model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/raw/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "architecture", "license" ], "notes": "The card says DeepSeek-V3.1 is post-trained on top of DeepSeek-V3.1-Base, lists both V3.1 and V3.1-Base as 671B total and 37B activated parameter models, describes a 128K context product target, states that model structure is the same as DeepSeek-V3 for local serving, and says model weights are MIT licensed. It also states V3.1 uses UE8M0 FP8 scale data format for model weights and activations." }, { "label": "DeepSeek V3.1 served config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/raw/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, hidden size 7168, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, BF16 torch dtype, FP8 e4m3 dynamic quantization with UE8M0 scales, and 163840 max position embeddings." }, { "label": "DeepSeek V3.1 Base config comparison", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base/raw/d3d4eafdc470de44bbf6f0a74f852eb522357be8/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "serving" ], "notes": "Manual comparison found matching checked architecture, routing, MLA, context, rope, dtype, and quantization fields between DeepSeek-V3.1 and DeepSeek-V3.1-Base. The base API also reports the same safetensors parameter split." }, { "label": "DeepSeek V3 official local inference code", "url": "https://github.com/deepseek-ai/DeepSeek-V3/blob/9b4e9788e4a3a731f7567338ed15d3ec549ce03b/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "ordinary_decode_scope" ], "notes": "Manual review found the default attn_impl absorb path. In that path MLA stores kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64, while the naive path would store expanded full K/V. The runtime supports scale_fmt for FP8 activation quantization and sets the default dtype to BF16 in the reference generator." }, { "label": "DeepSeek V3 official local inference config", "url": "https://github.com/deepseek-ai/DeepSeek-V3/blob/9b4e9788e4a3a731f7567338ed15d3ec549ce03b/inference/configs/config_671B.json", "source_type": "config", "supports": [ "compressed_state", "serving" ], "notes": "The official DeepSeek-V3 config records dtype fp8, 61 layers, 3 dense layers, 256 routed experts, 8 activated experts, 1 shared expert, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, and v_head_dim 128." }, { "label": "DeepSeek V3.1 safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/resolve/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensor bytes sum to 688.574839360 GB: F8_E4M3 680.571043840 GB, BF16 7.837573120 GB, and F32 0.166222400 GB. Main tensors in layers 0-60 plus embeddings, norm, and lm_head sum to 673.150611808 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic, including shared experts, sums to 19.082195296 GB. The auxiliary next-token-prediction layer 61 sums to 15.424227552 GB resident-only. The index metadata total_size field records 1369.062772 GB, but direct shard headers and the API dtype parameter counts agree on 688.574839360 GB of stored bytes." }, { "label": "DeepSeek V3.1 custom Hugging Face modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/raw/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "serving_caveat", "kv_adapter_boundary" ], "notes": "Manual review found that the bundled custom Transformers model expands MLA into key_states/value_states for the generic Transformers cache. The model card directs local serving users to the DeepSeek-V3 repo, so this profile follows the official absorb-MLA local inference path instead of the generic Transformers cache path." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned V3.1 and V3.1-Base configs, model card, official DeepSeek-V3 local inference config/code, custom HF modeling caveat, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without falling back to the stale generated full-KV estimate." }, { "id": "deepseek-ai--deepseek-v3-2-exp", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V3.2-Exp", "title": "DeepSeek V3.2 Exp FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 DeepSeek V3.2 Exp repo.", "model_family": "deepseek-v32-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V3.2-Exp-Base", "relation": "finetune", "source": "Hugging Face API cardData base_model fields, served config comparison, and direct shard-header audit", "config_compatible": true, "notes": "The Exp repo and Exp-Base config files are byte-identical at the reviewed commits. The Exp repo also has byte-identical config.json, inference/config_671B_v3.2.json, and inference/model.py files to the already audited deepseek-ai/DeepSeek-V3.2 release commit, and direct shard-header reads found the same resident and traffic byte decomposition." }, "architecture": { "canonical_architecture_id": "deepseek-v3-2", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 689.4711072, "main_resident_weight_gb": 674.032381632, "auxiliary_resident_weight_gb": 15.438725568, "fixed_weight_gb": 19.96396512, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Fixed traffic includes dense layers, shared experts, gates, norms, attention, DSA indexer weights, embeddings, and lm_head. Expected-distinct routing is applied to the 256 uniform routed experts across the main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Conservative cache allocation coefficient from official inference config/code: 61 layers * ((kv_lora_rank 512 FP8 bytes + 4 F32 scale values) + qk_rope_head_dim 64 BF16 bytes + (index_head_dim 128 FP8 bytes + 1 F32 scale value)) * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Bounds Engine v1 conservatively charges the same full-context MLA plus DSA index cache bytes per active read token. The V3.2 top-2048 DSA sparse read savings are not modeled by this linear formula." }, "notes": "DeepSeek V3.2 Exp uses MLA plus DeepSeek Sparse Attention. The official inference code keeps an MLA latent cache, a rope cache, and an FP8 indexer key cache. This profile uses the compressed_state adapter with audited coefficients rather than the older R1 MLA-only coefficient." }, "notes": "The official checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary layer from traffic but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_fp8_bf16_rope_dsa_fp8_index", "kv_store_bytes_per_scalar": 1, "kv_read_format": "conservative_full_context_mla_fp8_bf16_rope_dsa_fp8_index", "kv_read_bytes_per_scalar": 1, "runtime_format": "official-transformers-fp8-dsa-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization, sparse-attention kernel efficiency, and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 quantization with ue8m0 scales and BF16 model dtype. The official inference code states deployment uses FP8 KV cache and defines an FP8 DSA index cache with F32 scales." }, "evidence": [ { "label": "DeepSeek V3.2 Exp model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.2-Exp", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model", "serving", "total_params_b" ], "notes": "At commit 194c67e12b1b0d6df0ef373ddcf215bc84027409, the API reports a public non-gated MIT text-generation repo with FP8 and region:us tags, base_model deepseek-ai/DeepSeek-V3.2-Exp-Base, 229989 downloads, and safetensors parameters split across BF16: 3946184704, F8_E4M3: 681408069632, and F32: 42667040 tensors." }, { "label": "DeepSeek V3.2 Exp README", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp", "source_type": "model_card", "supports": [ "license", "base_model_proof", "dsa", "serving" ], "notes": "The README describes DeepSeek-V3.2-Exp as an experimental release introducing DeepSeek Sparse Attention for long-context efficiency. It documents local conversion/generation, SGLang launch, and vLLM day-0 support, and notes a November 17, 2025 inference demo code update fixing the RoPE layout in the indexer module." }, { "label": "DeepSeek V3.2 Exp config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV32ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 quantization with ue8m0 scales, kv_lora_rank 512, qk_rope_head_dim 64, index_head_dim 128, index_n_heads 64, index_topk 2048, and 163840 max position embeddings." }, { "label": "DeepSeek V3.2 Exp Base config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base/raw/b21ef155036dfbbd6d49013a345d31d5f5e38c89/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "serving" ], "notes": "At commit b21ef155036dfbbd6d49013a345d31d5f5e38c89, the base model config is byte-identical to the Exp repo config and exposes the same DeepseekV32ForCausalLM architecture, context length, routing fields, DSA index fields, and quantization fields used by this profile." }, { "label": "DeepSeek V3.2 Exp safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensors sum to 689.4711072 GB: F8_E4M3 681.408069632 GB, BF16 7.892369408 GB, and F32 0.170668160 GB. Main tensors in layers 0-60 plus embeddings/norm/head sum to 674.032381632 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic sums to 19.963965120 GB. The auxiliary next-token prediction layer 61 sums to 15.438725568 GB resident-only. The index metadata total_size is 1370.793842752 GB, so this profile uses direct tensor data_offsets as authoritative stored bytes." }, { "label": "DeepSeek V3.2 Exp official inference code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "dsa_index_cache" ], "notes": "Manual review found MLA caches for kv_lora_rank 512 and qk_rope_head_dim 64, plus an Indexer cache with index_head_dim 128 stored as FP8 and F32 scale values. The code states actual deployment uses FP8 KV cache and uses index_topk to select sparse attention positions. This file is byte-identical to the already audited deepseek-ai/DeepSeek-V3.2 inference/model.py at commit a7e62ac04ecb2c0a54d736dc46601c5606cf10a6." }, { "label": "DeepSeek V3.2 Exp official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/inference/config_671B_v3.2.json", "source_type": "config", "supports": [ "compressed_state", "serving" ], "notes": "The official inference config records dtype fp8, scale_fmt ue8m0, 61 layers, kv_lora_rank 512, qk_rope_head_dim 64, index_head_dim 128, index_n_heads 64, and index_topk 2048. It is byte-identical to the already audited deepseek-ai/DeepSeek-V3.2 inference config at commit a7e62ac04ecb2c0a54d736dc46601c5606cf10a6." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, config, base config, official inference code/config, safetensors index, direct range-read headers for all 163 shards, and byte-identity checks against the already audited DeepSeek V3.2 release files." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. The KV read formula is deliberately conservative until Bounds Engine supports a nonlinear DSA top-k sparse-read adapter." }, { "id": "deepseek-ai--deepseek-v3-2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V3.2", "title": "DeepSeek V3.2 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 DeepSeek V3.2 repo.", "model_family": "deepseek-v32-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V3.2-Exp-Base", "relation": "finetune", "source": "Hugging Face API cardData base_model and base_model_relation fields", "config_compatible": true, "notes": "The V3.2 and V3.2-Exp-Base configs expose the same DeepseekV32ForCausalLM shape and safetensors parameter split at the reviewed commits." }, "architecture": { "canonical_architecture_id": "deepseek-v3-2", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 689.4711072, "main_resident_weight_gb": 674.032381632, "auxiliary_resident_weight_gb": 15.438725568, "fixed_weight_gb": 19.96396512, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Fixed traffic includes dense layers, shared experts, gates, norms, attention, DSA indexer weights, embeddings, and lm_head. Expected-distinct routing is applied to the 256 uniform routed experts across the main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Conservative cache allocation coefficient from official inference config/code: 61 layers * ((kv_lora_rank 512 FP8 bytes + 4 F32 scale values) + qk_rope_head_dim 64 BF16 bytes + (index_head_dim 128 FP8 bytes + 1 F32 scale value)) * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Bounds Engine v1 conservatively charges the same full-context MLA plus DSA index cache bytes per active read token. The V3.2 top-2048 DSA sparse read savings are not modeled by this linear formula." }, "notes": "DeepSeek V3.2 uses MLA plus DeepSeek Sparse Attention. The official inference code keeps an MLA latent cache, a rope cache, and an FP8 indexer key cache. This profile uses the compressed_state adapter with audited coefficients rather than the older R1 MLA-only coefficient." }, "notes": "The official checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary layer from traffic but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_fp8_bf16_rope_dsa_fp8_index", "kv_store_bytes_per_scalar": 1, "kv_read_format": "conservative_full_context_mla_fp8_bf16_rope_dsa_fp8_index", "kv_read_bytes_per_scalar": 1, "runtime_format": "official-transformers-fp8-dsa-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization, sparse-attention kernel efficiency, and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 quantization with ue8m0 scales and BF16 model dtype. The official inference code states deployment uses FP8 KV cache and defines an FP8 DSA index cache with F32 scales." }, "evidence": [ { "label": "DeepSeek V3.2 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.2", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model", "serving", "total_params_b" ], "notes": "At commit a7e62ac04ecb2c0a54d736dc46601c5606cf10a6, the API reports an MIT text-generation repo with FP8 tag, base_model deepseek-ai/DeepSeek-V3.2-Exp-Base, and safetensors parameters split across BF16: 3946184704, F8_E4M3: 681408069632, and F32: 42667040 tensors." }, { "label": "DeepSeek V3.2 config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV32ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 quantization with ue8m0 scales, kv_lora_rank 512, qk_rope_head_dim 64, index_head_dim 128, index_n_heads 64, index_topk 2048, and 163840 max position embeddings." }, { "label": "DeepSeek V3.2 Exp Base config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base/raw/b21ef155036dfbbd6d49013a345d31d5f5e38c89/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "serving" ], "notes": "At commit b21ef155036dfbbd6d49013a345d31d5f5e38c89, the base model config exposes the same DeepseekV32ForCausalLM architecture, context length, routing fields, DSA index fields, and quantization fields used by this profile." }, { "label": "DeepSeek V3.2 safetensors index", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensors sum to 689.4711072 GB: F8_E4M3 681.408069632 GB, BF16 7.892369408 GB, and F32 0.17066816 GB. Main tensors in layers 0-60 plus embeddings/norm/head sum to 674.032381632 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic sums to 19.96396512 GB. The auxiliary next-token prediction layer 61 sums to 15.438725568 GB resident-only." }, { "label": "DeepSeek V3.2 official inference code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "dsa_index_cache" ], "notes": "Manual review found MLA caches for kv_lora_rank 512 and qk_rope_head_dim 64, plus an Indexer cache with index_head_dim 128 stored as FP8 and F32 scale values. The code states actual deployment uses FP8 KV cache and uses index_topk to select sparse attention positions." }, { "label": "DeepSeek V3.2 official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/inference/config_671B_v3.2.json", "source_type": "config", "supports": [ "compressed_state", "serving" ], "notes": "The official inference config records dtype fp8, scale_fmt ue8m0, 61 layers, kv_lora_rank 512, qk_rope_head_dim 64, index_head_dim 128, index_n_heads 64, and index_topk 2048." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from config, base config, HF API metadata, official inference code/config, safetensors index, and range-read safetensors header parameter/byte splits." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. The KV read formula is deliberately conservative until Bounds Engine supports a nonlinear DSA top-k sparse-read adapter." }, { "id": "deepseek-ai--deepseek-v3", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V3", "title": "DeepSeek V3 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 DeepSeek V3 repo.", "model_family": "deepseek-v3-moe", "architecture": { "canonical_architecture_id": "deepseek-v3", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 688.57483936, "main_resident_weight_gb": 673.150611808, "auxiliary_resident_weight_gb": 15.424227552, "fixed_weight_gb": 19.082195296, "routed_expert_weight_gb": 2.554954752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary decode through main layers 0-60, embeddings, norm, and lm_head, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 weights with BF16/F32 side tensors. Expected-distinct routing is applied to the 256 byte-uniform routed expert indexes across the 58 main MoE layers." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Official DeepSeek-V3 absorb-MLA cache coefficient: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent plus RoPE cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "The official DeepSeek-V3 local inference implementation uses the absorb MLA path by default, storing kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64. The generic Hugging Face custom modeling file expands to full K/V for the Transformers cache and is not the intended local-serving path." }, "notes": "The checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary next-token-prediction layer from per-token traffic but keeps it resident." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "official-deepseek-v3-fp8-absorb-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8/BF16/F32 checkpoint bytes from safetensors headers. FP8 dequantization, activation traffic, kernel efficiency, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records FP8 e4m3 quantization and BF16 model dtype. The official DeepSeek-V3 reference generator sets torch.set_default_dtype(torch.bfloat16), so the MLA latent/RoPE cache is charged as BF16." }, "evidence": [ { "label": "DeepSeek V3 model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit e815299b0bcbac849fa540c768ef21845365c9eb, the API reports a public text-generation repo with deepseek_v3, custom_code, fp8, and region:us tags, plus 1,113,089 downloads. The model card says DeepSeek-V3 has 671B main model parameters, 37B activated parameters, a 14B MTP module, FP8 weights only, MLA, DeepSeekMoE, 128K context, a model license with commercial use, and a warning that Hugging Face Transformers is not directly supported yet." }, { "label": "DeepSeek V3 served config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, hidden size 7168, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, FP8 e4m3 quantization, and 163840 max position embeddings." }, { "label": "DeepSeek V3 official local inference config and code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "ordinary_decode_scope" ], "notes": "Manual review of the pinned official inference/model.py and inference/configs/config_671B.json found the default attn_impl absorb path. In that path MLA stores kv_cache with kv_lora_rank 512 and pe_cache with qk_rope_head_dim 64, while the naive path would store expanded full K/V. The paired config_671B.json records the same 61-layer, 256-expert, 8-active-expert, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, FP8 model shape." }, { "label": "DeepSeek V3 safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/resolve/e815299b0bcbac849fa540c768ef21845365c9eb/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensor bytes sum to 688.574839360 GB: F8_E4M3 680.571043840 GB, BF16 7.837573120 GB, and F32 0.166222400 GB. Main tensors in layers 0-60 plus embeddings, norm, and lm_head sum to 673.150611808 GB. Main routed expert tensors for layers 3-60 sum to 654.068416512 GB, exactly 2.554954752 GB per expert index. Main fixed ordinary-decode traffic, including shared experts, sums to 19.082195296 GB. The auxiliary next-token-prediction layer 61 sums to 15.424227552 GB resident-only. The index metadata total_size field records 1369.062772 GB, but direct shard headers and the API dtype parameter counts agree on 688.574839360 GB of stored bytes." }, { "label": "DeepSeek V3 custom Hugging Face modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "serving_caveat", "kv_adapter_boundary" ], "notes": "Manual review found that the bundled custom Transformers model expands MLA into key_states/value_states for the generic Transformers cache. The model card states Hugging Face Transformers is not directly supported yet, so this profile follows the official DeepSeek-V3 local inference path instead of the generic Transformers cache path." }, { "label": "DeepSeek V3 model license", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/LICENSE-MODEL", "source_type": "model_card", "supports": [ "license" ], "notes": "The model license grants worldwide no-charge copyright and patent permissions subject to its terms, and the model card states DeepSeek-V3 series supports commercial use. Code files are separately MIT licensed." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the pinned served config, the model card, the official DeepSeek-V3 local inference config/code, the pinned custom HF modeling caveat, safetensors index, direct range-read shard header byte grouping, and the model license." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without falling back to the stale generated full-KV estimate." }, { "id": "deepseek-ai--deepseek-v4-flash", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V4-Flash", "title": "DeepSeek V4 Flash FP4/FP8", "summary": "Audited memory-side text-decode bounds profile for the official mixed FP4/FP8 DeepSeek V4 Flash repo.", "model_family": "deepseek-v4-flash-moe", "architecture": { "canonical_architecture_id": "deepseek-v4-flash", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 159.609485896, "main_resident_weight_gb": 154.95663638, "auxiliary_resident_weight_gb": 4.652849516, "fixed_weight_gb": 7.786897628, "routed_expert_weight_gb": 0.574881792, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp4_fp8_bf16_f32_i64", "traffic_scope": "ordinary decode through layers 0-42 plus norm.weight, head.weight, and top-level hc_head tensors, excluding resident-only embed.weight and mtp.0 tensors", "auxiliary_scope": "embed.weight and mtp.0 tensors are resident for the checkpoint but not swept for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "The model card reports logical 284B total / 13B active parameters. This profile uses exact stored tensor bytes for the mixed FP4/FP8 serving artifact because safetensors parameter counts understate logical FP4 expert parameters. Routed expert tensors are byte-uniform across 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 43, "kv_heads": 1, "head_dim": 512, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "All main layers keep a 128-token BF16 latent KV window in the official inference code." }, { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00688, "notes": "Allocation coefficient for 21 ratio-4 main compressed BF16 caches, 20 ratio-128 main compressed BF16 caches, and 21 ratio-4 indexer BF16 caches." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.001504, "notes": "Read coefficient for full ratio-4 indexer scoring cache plus ratio-128 compressed cache at the default read context; the capped ratio-4 selected main-KV read is represented as a fixed read term." }, "notes": "DeepSeek V4 Flash uses Compressed Sparse Attention and Heavily Compressed Attention. Bounds Engine v1 linearizes the compressed-cache pieces from the official inference code." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.01220608, "read_gb_per_output_token": 0.011010048, "state_formula": "Compressor kv_state/score_state fixed buffers across 21 ratio-4 layers, 20 ratio-128 layers, and 21 ratio-4 indexer layers; fixed read term charges 512 selected compressed main-KV slots for the 21 ratio-4 indexer layers", "notes": "The allocation is true fixed compressor state. The read term is a default-workload cap approximation for the sparse top-k selected main-KV read because Bounds Engine v1 does not yet model min(index_topk, context/compression) directly." } ], "notes": "At the default 100k allocated context and 32k read context, this profile charges 0.705842176 GB allocation and 0.064774144 GB read traffic per output token for the BF16 window/compressed/indexer cache path." }, "notes": "The official inference forward path constructs MTP modules but does not call them for ordinary text decode; MTP tensors remain resident-only in this profile." }, "serving": { "weight_format": "fp4_fp8_mixed", "weight_bytes_per_param": 1, "kv_store_format": "bf16_window_plus_compressed_bf16_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "sparse_bf16_window_compressed_kv_plus_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "official-deepseek-v4-fp4-fp8-csa-hca-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored FP4/FP8/BF16/F32/I64 safetensors bytes. FP4/FP8 dequantization, sparse-attention kernel efficiency, state writes, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 dynamic quantization plus expert_dtype fp4; the model card describes the artifact as FP4 + FP8 mixed with MoE expert params in FP4 and most other params in FP8." }, "evidence": [ { "label": "DeepSeek V4 Flash model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Flash", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "weight_format", "max_context_tokens" ], "notes": "At commit 60d8d70770c6776ff598c94bb586a859a38244f1, the API reports an MIT text-generation repo with deepseek_v4, fp8, and 8-bit tags. The card states DeepSeek-V4-Flash has 284B total parameters, 13B activated parameters, 1M context, and FP4 + FP8 mixed precision where MoE expert params use FP4 and most other params use FP8." }, { "label": "DeepSeek V4 Flash config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention", "serving" ], "notes": "The config records DeepseekV4ForCausalLM, 43 hidden layers, one MTP layer, 256 routed experts, 6 experts per token, 1 shared expert, hidden size 4096, 64 attention heads, 1 KV head, head_dim 512, 1048576 max position embeddings, sliding_window 128, index_head_dim 128, index_topk 512, FP8 quantization metadata, expert_dtype fp4, and compress_ratios with two uncompressed layers, 21 ratio-4 layers, 20 ratio-128 layers, and an uncalled MTP ratio entry." }, { "label": "DeepSeek V4 Flash safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "Safetensors headers were range-read across all 46 shards. Stored tensors sum to 159.609485896 GB: BF16 2.830518528 GB, I64 0.01861632 GB, F32 0.144672072 GB, F8_E8M0 8.858737664 GB, F8_E4M3 6.023020544 GB, and I8 141.733920768 GB. Ordinary swept tensors under layers plus norm, head, and top-level hc_head tensors sum to 154.95663638 GB. Resident-only embed.weight plus mtp.0 tensors sum to 4.652849516 GB. Routed expert tensors sum to 147.169738752 GB, or 0.574881792 GB per expert index. Fixed ordinary-decode traffic including shared experts sums to 7.786897628 GB." }, { "label": "DeepSeek V4 Flash official inference code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/inference/model.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "compressed_attention", "kv_adapter", "mtp_resident_only" ], "notes": "Manual review found Transformer.forward loops over the 43 main layers, applies norm and head, and does not call self.mtp in ordinary generation. Attention keeps a 128-token BF16 window, optional compressed BF16 cache slots according to compress_ratio, and an Indexer cache for ratio-4 layers. Ratio-4 indexer reads are capped by index_topk 512 for final sparse attention, while index scoring scans the compressed index cache." }, { "label": "DeepSeek V4 Flash official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/inference/config.json", "source_type": "config", "supports": [ "serving", "kv_adapter", "weight_format" ], "notes": "The official inference config records dtype fp8, expert_dtype fp4, scale_fmt ue8m0, 43 layers, 6 activated experts, 128-token window, index_topk 512, index_head_dim 128, and the same compression-ratio schedule used for the KV adapter arithmetic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, official inference config/code, safetensors index, and range-read safetensors header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems for the official checkpoint. The existing DS4/GGUF DeepSeek V4 profile remains the appropriate 128GB-oriented quantized artifact profile." }, { "id": "deepseek-ai--deepseek-v4-pro", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/DeepSeek-V4-Pro", "title": "DeepSeek V4 Pro FP4/FP8", "summary": "Audited memory-side text-decode bounds profile for the official mixed FP4/FP8 DeepSeek V4 Pro repo.", "model_family": "deepseek-v4-pro-moe", "architecture": { "canonical_architecture_id": "deepseek-v4-pro", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 864.704792696, "main_resident_weight_gb": 848.894846924, "auxiliary_resident_weight_gb": 15.809945772, "fixed_weight_gb": 26.840623052, "routed_expert_weight_gb": 2.140766208, "routed_experts": 384, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp4_fp8_bf16_f32_i64", "traffic_scope": "ordinary decode through layers 0-60 plus norm.weight, head.weight, and top-level hc_head tensors, excluding resident-only embed.weight and mtp.0 tensors", "auxiliary_scope": "embed.weight and mtp.0 tensors are resident for the checkpoint but not swept for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "The model card reports logical 1.6T total / 49B active parameters. This profile uses exact stored tensor bytes for the mixed FP4/FP8 serving artifact because logical FP4 expert parameters do not map one-to-one to stored bytes. Routed expert tensors are byte-uniform across 384 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 61, "kv_heads": 1, "head_dim": 512, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "All main layers keep a 128-token BF16 latent KV window in the official inference code." }, { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.009848, "notes": "Allocation coefficient for 30 ratio-4 main compressed BF16 caches, 31 ratio-128 main compressed BF16 caches, and 30 ratio-4 indexer BF16 caches." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.002168, "notes": "Read coefficient for full ratio-4 indexer scoring cache plus ratio-128 compressed cache at the default read context; the capped ratio-4 selected main-KV read is represented as a fixed read term." }, "notes": "DeepSeek V4 Pro uses Compressed Sparse Attention and Heavily Compressed Attention. Bounds Engine v1 linearizes the compressed-cache pieces from the official inference code." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.018710528, "read_gb_per_output_token": 0.03145728, "state_formula": "Compressor kv_state/score_state fixed buffers across 30 ratio-4 layers, 31 ratio-128 layers, and 30 ratio-4 indexer layers; fixed read term charges 1024 selected compressed main-KV slots for the 30 ratio-4 indexer layers", "notes": "The allocation is true fixed compressor state. The read term is a default-workload cap approximation for the sparse top-k selected main-KV read because Bounds Engine v1 does not yet model min(index_topk, context/compression) directly." } ], "notes": "At the default 100k allocated context and 32k read context, this profile charges 1.01150592 GB allocation and 0.108828672 GB read traffic per output token for the BF16 window/compressed/indexer cache path." }, "notes": "The official inference forward path constructs MTP modules but does not call them for ordinary text decode; MTP tensors remain resident-only in this profile." }, "serving": { "weight_format": "fp4_fp8_mixed", "weight_bytes_per_param": 1, "kv_store_format": "bf16_window_plus_compressed_bf16_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "sparse_bf16_window_compressed_kv_plus_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "official-deepseek-v4-pro-fp4-fp8-csa-hca-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored FP4/FP8/BF16/F32/I64 safetensors bytes. FP4/FP8 dequantization, sparse-attention kernel efficiency, state writes, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 dynamic quantization plus expert_dtype fp4; the model card describes the artifact as FP4 + FP8 mixed with MoE expert params in FP4 and most other params in FP8." }, "evidence": [ { "label": "DeepSeek V4 Pro model card and API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "weight_format", "max_context_tokens" ], "notes": "At commit b5968e9190ef611bbf34a7229255be88a0e937c1, the API reports an MIT text-generation repo with deepseek_v4, fp8, 8-bit, and region:us tags and 1140384 downloads. The card states DeepSeek-V4-Pro has 1.6T total parameters, 49B activated parameters, 1M context, and FP4 + FP8 mixed precision where MoE expert params use FP4 and most other params use FP8." }, { "label": "DeepSeek V4 Pro config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention", "serving" ], "notes": "The config records DeepseekV4ForCausalLM, 61 hidden layers, one MTP layer, 384 routed experts, 6 experts per token, 1 shared expert, hidden size 7168, 128 attention heads, 1 KV head, head_dim 512, 1048576 max position embeddings, sliding_window 128, index_head_dim 128, index_topk 1024, FP8 quantization metadata, expert_dtype fp4, and compress_ratios with 30 ratio-4 main layers, 31 ratio-128 main layers, and an uncalled MTP ratio entry." }, { "label": "DeepSeek V4 Pro safetensors index and shard headers", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/resolve/b5968e9190ef611bbf34a7229255be88a0e937c1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "Safetensors headers were range-read across all 64 shards. Stored tensors sum to 864.704792696 GB: BF16 5.633798656 GB, I64 0.01861632 GB, F32 0.351105656 GB, F8_E8M0 49.150268416 GB, F8_E4M3 23.169335296 GB, and I8 786.381668352 GB. Ordinary swept tensors under layers plus norm, head, and top-level hc_head tensors sum to 848.894846924 GB. Resident-only embed.weight plus mtp.0 tensors sum to 15.809945772 GB. Routed expert tensors sum to 822.054223872 GB, or 2.140766208 GB per expert index. Fixed ordinary-decode traffic including shared experts sums to 26.840623052 GB." }, { "label": "DeepSeek V4 Pro official inference code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/inference/model.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "compressed_attention", "kv_adapter", "mtp_resident_only" ], "notes": "Manual review found Transformer.forward loops over the 61 main layers, applies norm and head, and does not call self.mtp in ordinary generation. Attention keeps a 128-token BF16 window, optional compressed BF16 cache slots according to compress_ratio, and an Indexer cache for ratio-4 layers. Ratio-4 indexer reads are capped by index_topk 1024 for final sparse attention, while index scoring scans the compressed index cache." }, { "label": "DeepSeek V4 Pro official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/inference/config.json", "source_type": "config", "supports": [ "serving", "kv_adapter", "weight_format" ], "notes": "The official inference config records dtype fp8, expert_dtype fp4, scale_fmt ue8m0, 61 layers, 6 activated experts, 128-token window, index_topk 1024, index_head_dim 128, and the same compression-ratio schedule used for the KV adapter arithmetic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, official inference config/code, safetensors index, and range-read safetensors header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems for the official checkpoint. Smaller quantized/package variants should get their own profiles rather than inheriting this resident footprint." }, { "id": "deepseek-ai--deepseek-vl2-tiny", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "deepseek-ai/deepseek-vl2-tiny", "title": "DeepSeek VL2 Tiny BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 DeepSeek-VL2 Tiny vision-language MoE repo.", "model_family": "deepseek-vl2-tiny-moe", "architecture": { "canonical_architecture_id": "deepseek-vl2-tiny", "max_context_tokens": 4096, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 6.74100288, "main_resident_weight_gb": 5.53851136, "auxiliary_resident_weight_gb": 1.20249152, "fixed_weight_gb": 0.69409024, "routed_expert_weight_gb": 0.07569408, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through language.model.layers.*, language.model.norm.weight, and language.lm_head.weight, excluding the input embedding matrix and resident multimodal tensors", "auxiliary_scope": "language.model.embed_tokens.weight, vision tensors, projector tensors, image_newline, and view_seperator are resident for the multimodal package but are not swept as full matrices for each generated text token", "shared_expert_notes": "The language config records n_shared_experts 2. Shared expert tensors are included in fixed_weight_gb because the runtime adds shared_experts every MoE layer.", "notes": "Header-derived BF16 bytes are used instead of the generated metadata estimate. Routed expert tensors under language.model.layers.1-11.mlp.experts.* are byte-uniform across 64 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 12, "kv_heads": 10, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The embedded language_config records use_mla false, 12 decoder layers, hidden size 1280, 10 attention heads, and 10 KV heads. The pinned runtime selects the MHA/Llama attention branch when use_mla is false, so Bounds Engine v1 charges ordinary full-context BF16 key and value streams." }, "notes": "DeepseekVLV2ForCausalLM is multimodal. This profile models generated text-token decode after image embedding prefill, not vision encoder, tiling, or projector throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16_full_context_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16_full_context_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "deepseek-vl2-transformers-bf16-mha-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, image tiling, projector prefill, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16 and all safetensors payload bytes are BF16." }, "evidence": [ { "label": "DeepSeek VL2 Tiny API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-vl2-tiny", "source_type": "model_card", "supports": [ "repo", "revision", "downloads", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At repo SHA 66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa, the API reports a public, non-gated transformers image-text-to-text repo with DeepSeek model license metadata, region:us tag, 985805 downloads, and safetensors parameters BF16: 3370501440." }, { "label": "DeepSeek VL2 Tiny model card", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny/raw/66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa/README.md", "source_type": "model_card", "supports": [ "model_family", "pipeline", "license", "multimodal_scope", "serving" ], "notes": "The card identifies DeepSeek-VL2 as a MoE vision-language series, says VL2 Tiny is built on DeepSeekMoE-3B with 1.0B activated parameters, describes dynamic image tiling, shows BF16 loading, and demonstrates prepare_inputs_embeds followed by language_model.generate for response generation. The license section states the model license applies and supports commercial use." }, { "label": "DeepSeek VL2 Tiny config", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny/raw/66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "vision_projector" ], "notes": "The config records deepseek_vl_v2, bfloat16, language_config DeepseekV2ForCausalLM, use_mla false, 12 language layers, first_k_dense_replace 1, hidden size 1280, 10 attention heads, 10 KV heads, 4096 max position embeddings, 64 routed experts, 6 experts per token, 2 shared experts, vocab size 129280, SigLIP SO400M patch14 384 vision settings, and a 1280-dimension MLP projector." }, { "label": "DeepSeek VL2 Tiny VLM runtime", "url": "https://raw.githubusercontent.com/deepseek-ai/DeepSeek-VL2/ef9f91e2b6426536b83294c11742c27be66361b1/deepseek_vl2/models/modeling_deepseek_vl_v2.py", "source_type": "manual_review", "supports": [ "multimodal_scope", "ordinary_text_decode_scope" ], "notes": "Manual review of the pinned GitHub runtime found DeepseekVLV2ForCausalLM builds a vision encoder, projector, image sequence parameters, and a DeepseekV2ForCausalLM language model. prepare_inputs_embeds runs the vision/projector path before text generation, and forward delegates to self.language.forward." }, { "label": "DeepSeek VL2 Tiny language runtime", "url": "https://raw.githubusercontent.com/deepseek-ai/DeepSeek-VL2/ef9f91e2b6426536b83294c11742c27be66361b1/deepseek_vl2/models/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "moe_routing", "shared_experts", "kv_adapter" ], "notes": "Manual review found DeepseekV2DecoderLayer selects the MHA/Llama attention branch when use_mla is false, and uses DeepseekV2MoE for layers at or after first_k_dense_replace. DeepseekV2MoE routes to top-k experts and adds shared_experts every MoE layer." }, { "label": "DeepSeek VL2 Tiny safetensors index and header", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny/resolve/66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The single safetensors shard header was range-read directly. Stored tensor payloads sum to 6.741002880 GB, all BF16, matching index total_size. language tensors total 5.869468160 GB, vision tensors 0.856451200 GB, projector tensors 0.015078400 GB, and image_newline plus view_seperator 0.000005120 GB. language.model.embed_tokens.weight and language.lm_head.weight are separate 0.330956800 GB tensors, so the input embedding is resident-only while lm_head is swept for ordinary decode. Main resident ordinary text tensors excluding multimodal tensors and input embedding sum to 5.538511360 GB. Routed expert tensors sum to 4.844421120 GB, exactly 0.075694080 GB per expert index. Fixed ordinary-decode traffic including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head is 0.694090240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, pinned model card, pinned config, direct safetensors header range read, and pinned DeepSeek-VL2 GitHub runtime files." }, "notes": "This profile supersedes the generated metadata estimate, which undercounted active MoE traffic and treated the multimodal package as simple 3.3705B BF16 weights. It is for ordinary text decode after multimodal prefill." }, { "id": "doradus-ai--rnj-1-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Doradus-AI/RnJ-1-Instruct-FP8", "title": "Doradus RnJ-1 Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors FP8 RnJ-1 Instruct repo.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "DoradusAI/RnJ-1-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, served compressed-tensors config, quantization recipe, and direct safetensors header metadata", "config_compatible": true, "notes": "The model card and API metadata identify DoradusAI/RnJ-1-Instruct as the base model. The served config records Gemma3ForCausalLM text geometry and compressed-tensors FP8 dynamic W8A8 quantization with lm_head ignored. Direct safetensors headers confirm the expected FP8 layer tensors plus F32 embeddings, F32 lm_head, norms, and quantization scales." }, "architecture": { "canonical_architecture_id": "rnj-1-gemma3-8b-fp8", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.83734528, "swept_params_b": 8.312008704, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 11.995496448, "swept_weight_gb": 9.894150144, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "safetensors_header_stored_f8_e4m3_f32", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, quantization scale tensors, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 643 tensors totaling 8837345280 stored parameters and 11.995496448 GB payload bytes. model.embed_tokens.weight and lm_head.weight are stored separately as F32 tensors with shape [128256, 4096]. The config records tie_word_embeddings true, but compressed-tensors quantization ignored lm_head and the checkpoint stores a separate F32 output projection, so only the input embedding tensor is resident-only for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 8 KV heads, 128 head dimension, max_position_embeddings 32768, and sliding_window 32768. Because the window equals the supported context, the profile charges full-context K and V streams for all language layers." }, "notes": "Dense Gemma3ForCausalLM text-generation profile using the served compressed-tensors config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.3573642386868472, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-compressed-tensors-fp8-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and F32 embeddings, lm_head, norms, and scale tensors from safetensors headers. Dynamic FP8 activation quantization, dequantization, compute throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records compressed-tensors quant_method with float-quantized 8-bit weights and token-dynamic 8-bit activations, ignored lm_head, and no quantized KV cache scheme. KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Doradus RnJ-1 Instruct FP8 HF API metadata", "url": "https://huggingface.co/api/models/Doradus-AI/RnJ-1-Instruct-FP8", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The live HF API response at commit 3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d records a public Gemma-license text-generation repo with transformers, safetensors, compressed-tensors, region:us, 527182 downloads, base_model DoradusAI/RnJ-1-Instruct, safetensors parameters F32 1052717056, F8_E4M3 7784628224, and total 8837345280." }, { "label": "Doradus RnJ-1 Instruct FP8 model card", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "serving", "max_context_tokens" ], "notes": "The card says this is an FP8 dynamic W8A8 quantized version of DoradusAI/RnJ-1-Instruct created with llmcompressor, with FP8 E4M3 weights and activations, lm_head kept in BF16/F32, vLLM and SGLang support, 32 layers, 32 attention heads, 8 KV heads, 4096 hidden size, 16384 intermediate size, 32768 max context, 128256 vocabulary, and roughly 8B parameters." }, { "label": "Doradus RnJ-1 Instruct FP8 config", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/raw/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "serving", "kv_adapter", "tie_word_embeddings" ], "notes": "The config records Gemma3ForCausalLM, model_type gemma3_text, tie_word_embeddings true, hidden size 4096, intermediate size 16384, 32 layers, 32 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 32768, sliding_window 32768, rope_theta 10000, vocab size 128256, and compressed-tensors quantization with FP8 dynamic activations, FP8 weights, ignored lm_head, kv_cache_scheme null, and version 0.12.2." }, { "label": "Doradus RnJ-1 Instruct FP8 safetensors index and linked-object HEAD checks", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/resolve/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "total_params_b", "weight_format" ], "notes": "The safetensors index records total_parameters 8837345280 and total_size 11995496448 across three shards. HEAD checks found linked shard sizes 4.956478440 GB, 4.937745728 GB, and 2.101346432 GB; the difference from tensor payload bytes is the safetensors header overhead." }, { "label": "Doradus RnJ-1 Instruct FP8 safetensors headers", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/tree/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "weight_format" ], "notes": "Range-reads of the three safetensors shard headers found 643 tensors totaling 11.995496448 GB: F8_E4M3 tensors total 7.784628224 GB and F32 tensors total 4.210868224 GB. model.embed_tokens.weight is F32 [128256, 4096] and contributes 2.101346304 GB resident-only for ordinary decode. lm_head.weight is a separate F32 [128256, 4096] tensor and remains in swept decode traffic. Layer tensors, quantization scale tensors, model.norm.weight, and lm_head.weight total 9.894150144 GB." }, { "label": "Doradus RnJ-1 Instruct FP8 quantization recipe", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/raw/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d/recipe.yaml", "source_type": "config", "supports": [ "serving", "weight_format" ], "notes": "The recipe targets Linear modules with FP8_DYNAMIC and ignores lm_head, matching the served compressed-tensors config and safetensors header split." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, quantization recipe, direct safetensors index and linked-object checks, and direct safetensors header byte grouping." }, "notes": "This profile intentionally does not assume FP8 KV cache or runtime speedups; it models memory-side text decode with BF16 KV and exact stored weight bytes." }, { "id": "douyamv--gemma-4-31b-jang-4m-crack-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF", "title": "Douyamv Gemma 4 31B JANG CRACK GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended Q4_K_M GGUF artifact of Gemma 4 31B JANG CRACK.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "derived_package", "source": "Hugging Face model card/API GGUF metadata, Google Gemma 4 31B base profile, linked-object sizes, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card identifies google/gemma-4-31B-it as the base model and says this package applies CRACK abliteration plus JANG v2 mixed-precision conversion before emitting standard GGUF quantizations. The selected Q4_K_M GGUF header records the same Gemma 4 31B dense text geometry as the audited Google profile: 60 layers, 32 attention heads, ten full-attention layers with four KV heads and 512 key/value length, 50 sliding-window layers with 16 KV heads and 256 key/value length, 1024-token sliding window, 262144 context, and tied output represented by token_embd.weight with no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 18.687057344, "swept_weight_gb": 18.671229168, "auxiliary_resident_weight_gb": 0.015828176, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31b-jang-crack-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; the selected repo does not include separate mmproj, vision, audio, or MTP sidecar files", "notes": "The profile targets Q4_K_M because the model card describes it as the best size/quality balance and uses it in the llama.cpp example command. Header tensor spans total 18.671229168 GB, while the linked file size is 18.687057344 GB. The HF API gguf.totalFileSize points at the smaller Q3_K_M artifact, so that mismatch is recorded but not used as the practical selected artifact." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so this profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact. The repo card mentions Gemma 4 multimodal features, but this repo's available GGUF files are main text artifacts with no separate multimodal projector sidecars." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6087515705734181, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and any multimodal prefill path are outside Bounds Engine v1.", "notes": "The selected artifact uses a Q4_K_M mixed tensor layout: tensor payloads split into Q4_K, Q6_K, and F32 classes. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Douyamv Gemma 4 31B JANG CRACK GGUF API metadata", "url": "https://huggingface.co/api/models/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "api_selected_artifact" ], "notes": "At commit ed3a8120d4da414573688d0130f061d3b4130372, the API records a public non-gated Gemma-license GGUF text-generation repo with base_model google/gemma-4-31B-it, gemma4, quantized, 31b, endpoints_compatible, region:us, 298877 downloads, GGUF architecture gemma4, context_length 262144, gguf.total 30697345596, and gguf.totalFileSize 15287102912. The API totalFileSize matches Q3_K_M, but the card recommends and demonstrates Q4_K_M." }, { "label": "Douyamv Gemma 4 31B JANG CRACK GGUF model card", "url": "https://huggingface.co/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "runtime_format" ], "notes": "The card identifies google/gemma-4-31B-it as the base model, describes CRACK abliteration and JANG v2 mixed-precision conversion, lists Q3_K_M and Q4_K_M as available quantizations, labels Q4_K_M as the best size/quality balance, and uses gemma-4-31b-jang-crack-Q4_K_M.gguf in the llama.cpp example." }, { "label": "Google Gemma 4 31B IT audited base profile", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records Gemma4ForConditionalGeneration, 60 text layers, ten full-attention layers, 50 sliding-window layers, 1024-token sliding window, attention_k_eq_v true, 4 global KV heads, 16 local KV heads, 512 global key/value length, 256 local key/value length, tied embeddings, and 262144 max position embeddings." }, { "label": "Douyamv Gemma 4 31B GGUF linked-object HEAD checks", "url": "https://huggingface.co/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF/tree/ed3a8120d4da414573688d0130f061d3b4130372", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "api_selected_artifact" ], "notes": "HEAD checks found Q3_K_M 15.287102912 GB, Q4_K_M 18.687057344 GB, and sharded Q8_0 parts totaling 32.635670976 GB. The API gguf.totalFileSize matches Q3_K_M, while the card-recommended Q4_K_M artifact is selected for this profile." }, { "label": "Douyamv Gemma 4 31B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF/resolve/ed3a8120d4da414573688d0130f061d3b4130372/gemma-4-31b-jang-crack-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 44 metadata entries and 833 tensors. The linked file is 18.687057344 GB, with tensor payloads starting at byte 15826496. Tensor spans sum to 18.671229168 GB across 30697345596 logical elements: token_embd.weight 1.156055040 GB, blk.* tensors 17.515151600 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_K 14.213283840 GB, Q6_K 4.452618240 GB, and F32 0.005327088 GB. Metadata/tokenizer/header/file overhead accounts for 0.015828176 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, layer KV head array with ten full layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for full layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight, mmproj, vision, audio, or MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited Google Gemma 4 31B IT base profile, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the card-recommended Q4_K_M main text GGUF artifact. Do not infer the API-matched Q3_K_M, sharded Q8_0, multimodal projector residency, or sidecar traffic unless those artifacts are explicitly selected and audited." }, { "id": "empero-ai--qwythos-9b-claude-mythos-5-1m-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF", "title": "Empero Qwythos 9B Claude Mythos 5 1M GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended Q4_K_M GGUF artifact of Qwythos 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "empero-ai/Qwythos-9B-Claude-Mythos-5-1M", "relation": "quantized", "source": "Hugging Face model card metadata, base-model config, upstream Qwen config comparison, and GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo records empero-ai/Qwythos-9B-Claude-Mythos-5-1M as its base model. That base repo records Qwen/Qwen3.5-9B as its base and keeps the Qwen3.5 9B layer geometry while changing the served context from 262144 to 1048576 tokens through YaRN. The selected GGUF header records the same Qwen3.5 9B geometry and 1048576-token context." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b-yarn-1m", "max_context_tokens": 1048576, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.953803264, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.01711872, "resident_weight_gb": 5.629109152, "swept_weight_gb": 5.046011904, "auxiliary_resident_weight_gb": 0.583097248, "resident_parameter_scope": "logical GGUF parameters from HF API and selected Q4_K_M header tensor shapes", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected normal text Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept for ordinary text decode; MTP and mmproj sidecar GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf because the repo card calls Q4_K_M the recommended default for normal text weights. The HF API gguf.totalFileSize points at the BF16 conversion base, so this profile records that evidence but does not use BF16 as the repo-level ordinary-serving default. The selected Q4_K_M linked file is 5.629109152 GB. Header tensor spans total 5.618141184 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.010967968 GB. The main GGUF contains output.weight, token_embd.weight, blk.* tensors, and output_norm.weight; it has no mmproj, vision, audio, MTP, or rope_freqs tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 layers with every fourth layer using full attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the card-recommended normal text Q4_K_M GGUF artifact after any multimodal prefill. It does not include the MTP variants or the separate vision projector sidecars." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6286835868543827, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and MTP speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected normal text artifact uses a mixed Q4_K_M layout: tensor spans split into 3.7748736 GB Q4_K, 1.83902208 GB Q6_K, and 0.004245504 GB F32. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Empero Qwythos 9B GGUF HF API metadata", "url": "https://huggingface.co/api/models/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit a7ab0f6ee807c165e8374e4906773ca39f5fdff3 records a public Apache-2.0 image-text-to-text GGUF repo with base_model empero-ai/Qwythos-9B-Claude-Mythos-5-1M, region:us, 1250562 downloads, GGUF architecture qwen35, 1048576 context length, gguf.total 8953803264, and gguf.totalFileSize 17920697248. The API totalFileSize matches the BF16 conversion base, while the model card recommends Q4_K_M as the default normal text artifact." }, { "label": "Empero Qwythos 9B GGUF model card", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The card describes fixed v3 GGUFs for the Qwythos 9B base model, says Q4_K_M is the recommended default for normal text weights, lists BF16 as the full precision conversion base, lists separate MTP variants for draft speculation, and lists separate mmproj files required for image input." }, { "label": "Empero Qwythos 9B base config", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M/raw/763f72fc2c3b186e977adcbaac0c18128f182166/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 1048576 max position embeddings, and a resident vision config. config.json.pre_yarn in the same repo records the upstream 262144-token context before the YaRN change." }, { "label": "Qwen3.5 9B upstream config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison against Qwen/Qwen3.5-9B found the same Qwen3.5 9B text geometry as Qwythos except for the served context extension from 262144 to 1048576 tokens." }, { "label": "Empero Qwythos 9B GGUF linked-object HEAD checks", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF/tree/a7ab0f6ee807c165e8374e4906773ca39f5fdff3", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found normal text files Q4_K_M 5.629109152 GB, Q5_K_M 6.467969952 GB, Q6_K 7.359259552 GB, Q8_0 9.527501728 GB, and BF16 17.920697248 GB. Separate MTP text variants range from 5.887668064 GB to 18.40732144 GB, and separate mmproj files are 0.918165472 GB and 0.91816608 GB. The selected profile uses Q4_K_M because the card recommends it as the default normal text artifact." }, { "label": "Empero Qwythos 9B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF/resolve/a7ab0f6ee807c165e8374e4906773ca39f5fdff3/Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 35 metadata entries and 427 tensors. The linked file is 5.629109152 GB. Tensor spans sum to 5.618141184 GB: output.weight 0.8343552 GB, token_embd.weight 0.57212928 GB, blk.* tensors 4.21164032 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.010967968 GB. Tensor spans split into Q4_K 3.7748736 GB, Q6_K 1.83902208 GB, and F32 0.004245504 GB. The header records qwen35.block_count 32, context_length 1048576, YaRN scaling factor 4 from original context 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no mmproj/vision/audio/MTP tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, base and upstream configs, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the Empero Qwythos normal text Q4_K_M GGUF artifact. Do not infer MTP speculative sidecar traffic, BF16 conversion-base residency, or multimodal projector residency unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "ggml-org--gemma-4-12b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ggml-org/gemma-4-12B-it-GGUF", "title": "ggml-org Gemma 4 12B IT GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected ggml-org BF16 GGUF artifact of Gemma 4 12B IT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face base_model metadata, README, API GGUF metadata, GGUF header metadata, and existing Google Gemma 4 12B IT profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it. The selected GGUF header records the same Gemma 4 12B text geometry and hybrid local/global attention pattern as the audited Google base profile." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 23.832064928, "swept_weight_gb": 23.816242688, "auxiliary_resident_weight_gb": 0.01582224, "resident_parameter_scope": "selected GGUF linked file size for gemma-4-12B-it-bf16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 23.816242688 GB, while the linked file size is 23.832064928 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact. Image, video, audio, and projector sidecar work requires separate workload profiles if explicitly loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the ggml-org BF16 GGUF because HF API gguf.totalFileSize matches gemma-4-12B-it-bf16.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param records the BF16 tensor storage format." }, "evidence": [ { "label": "ggml-org Gemma 4 12B IT GGUF API metadata", "url": "https://huggingface.co/api/models/ggml-org/gemma-4-12B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The API response at commit 44ee90c4b61e888ac5b318a54ec7a94df61e9cd7 records base_model google/gemma-4-12B-it, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, gguf.totalFileSize 23832064928, 286749 downloads, endpoints_compatible, conversational, and region:us." }, { "label": "ggml-org Gemma 4 12B IT GGUF README", "url": "https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format" ], "notes": "The README records base_model google/gemma-4-12B-it and recommends running with llama-server -hf ggml-org/gemma-4-12B-it-GGUF." }, { "label": "Google Gemma 4 12B IT audited profile", "url": "https://huggingface.co/google/gemma-4-12B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "attention_pattern", "kv_adapter", "tie_word_embeddings", "max_context_tokens" ], "notes": "The existing audited Google profile records 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, tied embeddings, and 262144 max positions." }, { "label": "Google Gemma 4 12B IT config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "ggml-org Gemma 4 12B IT GGUF linked sizes", "url": "https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF/tree/44ee90c4b61e888ac5b318a54ec7a94df61e9cd7", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-12B-it-bf16.gguf is 23832064928 bytes, exactly matching API gguf.totalFileSize. Sibling files are gemma-4-12B-it-Q4_K_M.gguf at 7381382048 bytes, gemma-4-12B-it-Q8_0.gguf at 12669645728 bytes, mmproj-gemma-4-12B-it-Q8_0.gguf at 158987584 bytes, and mmproj-gemma-4-12B-it-bf16.gguf at 175115584 bytes." }, { "label": "ggml-org Gemma 4 12B IT BF16 GGUF range-read tensor index", "url": "https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF/resolve/44ee90c4b61e888ac5b318a54ec7a94df61e9cd7/gemma-4-12B-it-bf16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 49 metadata entries and 667 tensors. The selected file is 23.832064928 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15822240. Tensor spans total 23.816242688 GB across 11907350576 logical elements: token_embd.weight 2.013265920 GB, blk.* tensors 21.802960384 GB, output_norm.weight 0.000015360 GB, and rope_freqs.weight 0.000001024 GB. Actual tensor bytes are 23.816241344 GB, with 1344 bytes of tensor-alignment padding. Stored tensor bytes split into BF16 23.813160960 GB and F32 0.003080384 GB; tensor spans split into BF16 23.813160960 GB and F32 0.003081728 GB. Metadata/tokenizer/header/file overhead accounts for 0.015822240 GB. The header records general.license apache-2.0, gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, README, linked-object HEAD checks for all GGUF files, the existing Google base profile/config, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the ggml-org API-selected BF16 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "ggml-org--gemma-4-26b-a4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ggml-org/gemma-4-26B-A4B-it-GGUF", "title": "ggml-org Gemma 4 26B A4B IT GGUF BF16 MoE", "summary": "Audited memory-side bounds profile for the API-selected ggml-org BF16 GGUF artifact of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face base_model metadata, README, API GGUF metadata, GGUF header metadata, and existing Google Gemma 4 26B A4B IT profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-26B-A4B-it. The selected GGUF header records the same Gemma 4 26B A4B text architecture and hybrid local/global attention pattern as the audited Google base profile." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 50.505134624, "main_resident_weight_gb": 50.489312376, "auxiliary_resident_weight_gb": 0.015822248, "fixed_weight_gb": 4.813341816, "routed_expert_weight_gb": 0.35684352, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for gemma-4-26B-A4B-it-bf16.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected BF16 main GGUF artifact", "auxiliary_scope": "GGUF metadata, header, and alignment padding are resident in the selected artifact but not swept as model tensors; separate mmproj GGUF files in the repo are not included unless explicitly loaded for multimodal workloads", "shared_expert_notes": "The GGUF metadata records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects gemma-4-26B-A4B-it-bf16.gguf. A GGUF v3 range-read found 658 tensors and 47 metadata entries. Tensor spans total 50.489312376 GB, while the linked file is 50.505134624 GB. Routed expert tensors total 45.675970560 GB and divide exactly by 128 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. The audited Google Gemma 4 profile records attention_k_eq_v true, so full-attention layers charge one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format." }, "notes": "This profile targets the API-selected BF16 main text GGUF artifact. The Q4_K_M, Q8_0, and mmproj sidecar files are separate artifacts and should get separate profiles if they are selected." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and non-selected quantized GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the ggml-org BF16 GGUF. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "ggml-org Gemma 4 26B A4B IT GGUF API metadata", "url": "https://huggingface.co/api/models/ggml-org/gemma-4-26B-A4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The API response at commit ae4d537a6345467d1c86bb5cc0d4505ff3ebe0f3 records base_model google/gemma-4-26B-A4B-it, GGUF architecture gemma4, 262144 context length, gguf.total 25233142046, gguf.totalFileSize 50505134624, 433222 downloads, endpoints_compatible, conversational, and region:us." }, { "label": "ggml-org Gemma 4 26B A4B IT GGUF README", "url": "https://huggingface.co/ggml-org/gemma-4-26B-A4B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format" ], "notes": "The README records base_model google/gemma-4-26B-A4B-it and recommends running with llama-server -hf ggml-org/gemma-4-26B-A4B-it-GGUF." }, { "label": "Google Gemma 4 26B A4B IT audited profile", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "kv_adapter" ], "notes": "The existing audited Google profile records 30 layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, attention_k_eq_v true, one shared expert, and 262144 max positions." }, { "label": "ggml-org Gemma 4 26B A4B IT BF16 GGUF linked-object and tensor-index range read", "url": "https://huggingface.co/ggml-org/gemma-4-26B-A4B-it-GGUF/resolve/ae4d537a6345467d1c86bb5cc0d4505ff3ebe0f3/gemma-4-26B-A4B-it-bf16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "HF linked-object metadata reports 50.505134624 GB for gemma-4-26B-A4B-it-bf16.gguf, matching API gguf.totalFileSize. A 64MB range-read of the GGUF v3 header found 658 tensors and 47 metadata entries. Tensor spans sum to 50.489312376 GB; metadata/header/alignment padding accounts for 0.015822248 GB. BF16 tensors account for 50.443255808 GB and F32 tensors for 0.046056568 GB. Fixed non-expert tensor spans total 4.813341816 GB. Routed expert tensors total 45.675970560 GB across 30 layers and 128 expert indexes, or 0.356843520 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, and the sliding-window pattern." }, { "label": "ggml-org Gemma 4 26B A4B IT GGUF linked sizes", "url": "https://huggingface.co/ggml-org/gemma-4-26B-A4B-it-GGUF/tree/ae4d537a6345467d1c86bb5cc0d4505ff3ebe0f3", "source_type": "manual_review", "supports": [ "selected_artifact" ], "notes": "HEAD checks found Q4_K_M 16.796015136 GB, Q8_0 26.859858464 GB, BF16 50.505134624 GB, mmproj Q8_0 0.806408000 GB, and mmproj BF16 1.194827840 GB. The profile targets BF16 because it is the file size reported by the API GGUF metadata." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, README, linked-object HEAD checks for all GGUF files, the existing Google base profile, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the ggml-org API-selected BF16 GGUF text artifact. Do not infer it from the Unsloth MXFP4_MOE GGUF profile, the Google BF16 safetensors profile, or the separate mmproj files." }, { "id": "ggml-org--gpt-oss-120b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ggml-org/gpt-oss-120b-GGUF", "title": "ggml-org gpt-oss 120B GGUF MXFP4_MOE", "summary": "Audited memory-side bounds profile for the ggml-org split MXFP4_MOE GGUF package of gpt-oss-120b.", "model_family": "gpt-oss-moe", "base_model_proof": { "base_model": "openai/gpt-oss-120b", "relation": "quantized", "source": "Hugging Face API base_model metadata, model card metadata, OpenAI base metadata, and direct split GGUF header metadata", "config_compatible": true, "notes": "The repo API and model card identify this package as a GGUF derivative of openai/gpt-oss-120b. The split GGUF metadata records the same checked gpt-oss geometry: 36 layers, 128 experts, 4 experts per token, 8 KV heads, 64 key/value length, 128-token sliding attention window, and 131072 context length." }, "architecture": { "canonical_architecture_id": "gpt-oss-120b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 63.387346464, "main_resident_weight_gb": 62.758994688, "auxiliary_resident_weight_gb": 0.628351776, "fixed_weight_gb": 1.685668608, "routed_expert_weight_gb": 0.47713536, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected three-part gpt-oss-120b-mxfp4 GGUF split linked file sizes including metadata-only first split, tensor spans, tokenizer, and GGUF overhead", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all blk.0 through blk.35 non-routed tensors, routers, expert biases, and expected-distinct routed expert weight groups from the selected MXFP4_MOE split package", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident across the selected split package but not swept for each ordinary text decode token", "notes": "The selected split package stores routed expert weights as MXFP4 tensors, output/token embeddings and attention tensors as Q8_0, and norms, routers, attention biases, expert biases, and metadata tensors as F32. The routed expert weight and bias tensors are byte-uniform across the 128 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 18, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "Odd-numbered layers use full_attention in the gpt-oss config family." }, { "kind": "sliding_window", "layers": 18, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "kv_scalar_multiplier": 2, "notes": "Even-numbered layers use sliding_attention with a 128-token window." } ], "notes": "The split GGUF metadata records the same 36-layer gpt-oss geometry used by the OpenAI base package; llama.cpp serving charges FP16 KV unless a separate runtime configuration overrides it." }, "notes": "This profile models ordinary text decode for the repo-level ggml-org split MXFP4_MOE GGUF package. It is not interchangeable with Unsloth's API-selected F16 GGUF artifact because the side tensors are Q8_0 rather than F16." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5425644442676337, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-mxfp4-moe-q8-0-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked split sizes for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and speculative behavior are outside Bounds Engine v1.", "notes": "The selected package is the three-part gpt-oss-120b-mxfp4 split published by ggml-org. GGUF file_type 38 marks mostly MXFP4 MoE; direct headers show Q8_0 side tensors and MXFP4 routed experts." }, "evidence": [ { "label": "ggml-org gpt-oss 120B GGUF API metadata", "url": "https://huggingface.co/api/models/ggml-org/gpt-oss-120b-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "max_context_tokens", "selected_artifact", "total_params_b" ], "notes": "The live HF API response at commit d932fcea62f83e088d8f076a2cd2d7eb02dfa682 records a public non-gated GGUF repo with base_model openai/gpt-oss-120b, current downloads 69411, region:us, GGUF architecture gpt-oss, context length 131072, gguf.total 116829156672, and gguf.totalFileSize 63387346464. The catalog row keeps the original qualifying scrape count 116699 until the over-100k working set is regenerated." }, { "label": "ggml-org gpt-oss 120B GGUF model card", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card metadata lists base_model openai/gpt-oss-120b. The visible card links to the llama.cpp guide and gives a llama-server quick-start using the ggml-org repo." }, { "label": "OpenAI gpt-oss 120B base metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-120b", "source_type": "model_card", "supports": [ "base_model_proof", "model_family", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving" ], "notes": "The base repo records Apache-2.0 text-generation metadata, GptOssForCausalLM/gpt_oss config, 36 layers, 128 local experts, 4 experts per token, alternating sliding/full attention, 128-token sliding windows, 8 KV heads, 64 head dimension, 131072 max position embeddings, and MXFP4 quantization in the official package." }, { "label": "ggml-org gpt-oss 120B split GGUF HEAD and header audit", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF/tree/d932fcea62f83e088d8f076a2cd2d7eb02dfa682", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "HEAD checks of the three selected split files found sizes 0.012980384 GB, 31.738487200 GB, and 31.635878880 GB, totaling 63.387346464 GB. Direct GGUF v3 range reads found a metadata-only first split plus 687 tensors in splits 2 and 3. Tensor spans sum to 63.374323968 GB and GGUF metadata/header/file overhead is 0.013022496 GB. Tensor spans split into MXFP4 60.914073600 GB, Q8_0 2.245893120 GB, and F32 0.214357248 GB. token_embd.weight is 0.615329280 GB and resident-only; output.weight is 0.615329280 GB and swept. Routed expert weights plus expert biases sum to 61.073326080 GB, or 0.477135360 GB per expert index. Fixed ordinary text traffic, including output.weight, attention tensors, routers, norms, and output_norm.weight, sums to 1.685668608 GB. The first split header records gpt-oss block_count 36, context_length 131072, expert_count 128, expert_used_count 4, attention.head_count 64, attention.head_count_kv 8, key/value length 64, 128-token sliding window, split.count 3, and split.tensors.count 687." }, { "label": "Selected split GGUF artifact", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF/resolve/d932fcea62f83e088d8f076a2cd2d7eb02dfa682/gpt-oss-120b-mxfp4-00001-of-00003.gguf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "serving" ], "notes": "The selected repo package consists of gpt-oss-120b-mxfp4-00001-of-00003.gguf, gpt-oss-120b-mxfp4-00002-of-00003.gguf, and gpt-oss-120b-mxfp4-00003-of-00003.gguf. The first split stores full metadata and no tensor payload; the second and third splits store the 687 tensor entries." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, model card metadata, OpenAI base metadata, split linked file HEAD checks, and direct range reads of all three selected GGUF split headers." }, "notes": "Use this profile for the ggml-org split MXFP4_MOE GGUF package in ordinary text-decode bounds. Do not substitute the Unsloth F16 GGUF profile or the official safetensors profile; their side-tensor residency and ordinary decode traffic differ." }, { "id": "ggml-org--smolvlm-500m-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ggml-org/SmolVLM-500M-Instruct-GGUF", "title": "ggml-org SmolVLM 500M Instruct GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 main GGUF artifact of SmolVLM 500M Instruct.", "model_family": "smolvlm-dense-multimodal-gguf", "base_model_proof": { "base_model": "HuggingFaceTB/SmolVLM-500M-Instruct", "relation": "quantized", "source": "Hugging Face base_model metadata, README, base config, API GGUF metadata, selected main GGUF header metadata, and mmproj sidecar review", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of HuggingFaceTB/SmolVLM-500M-Instruct. The selected main GGUF header records the same text geometry as the base config: VLlama3ForCausalLM text decoder, 32 layers, 15 attention heads, 5 KV heads, 64 key/value head dimension, 960 hidden size, 2560 intermediate size, 8192 context, and untied output embeddings. The selected main artifact excludes the separate mmproj vision/projector sidecar files, which are recorded as out of scope for this ordinary text-decode profile unless a workload explicitly loads them." }, "architecture": { "canonical_architecture_id": "smolvlm-500m-instruct-text", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.4092528, "swept_params_b": 0.361944, "auxiliary_resident_params_b": 0.0473088, "resident_weight_gb": 0.820422912, "swept_weight_gb": 0.7240128, "auxiliary_resident_weight_gb": 0.096410112, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for SmolVLM-500M-Instruct-f16.gguf", "swept_parameter_scope": "ordinary text decode charges output.weight, blk.* tensors, and output_norm.weight in the selected main text artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected main artifact but not swept as decode weights; separate mmproj GGUF files are not included unless explicitly loaded for a multimodal workload", "notes": "The selected main F16 GGUF stores a separate output.weight tensor, so token_embd.weight is resident-only for ordinary decode. Header tensor spans total 0.818630400 GB while the linked main file is 0.820422912 GB. The main file contains output.weight, token_embd.weight, blk.* tensors, and output_norm.weight. It has no mmproj, vision, audio, MTP, or draft tensor." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 5, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The base text config and selected main GGUF metadata record full-context Llama-style text attention with 32 layers, 5 KV heads, and 64-dimensional key/value heads." }, "notes": "SmolVLM is multimodal, but this profile models ordinary cached text decode for the selected main GGUF artifact. Image prefill, mmproj execution, and mmproj sidecar residency require a separate workload/profile scope." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0046849086921337, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for swept weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is SmolVLM-500M-Instruct-f16.gguf because HF API gguf.totalFileSize matches that linked object. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes over API GGUF logical tensor parameters, including GGUF overhead." }, "evidence": [ { "label": "ggml-org SmolVLM 500M Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-500M-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit 72e986006ef53e37cdd3f6d4241c90b0f01df376, the API records a public non-gated Apache-2.0 GGUF repo with base_model HuggingFaceTB/SmolVLM-500M-Instruct, endpoints_compatible, region:us, 178205 downloads, GGUF architecture llama, 8192 context length, gguf.total 409252800, and gguf.totalFileSize 820422912." }, { "label": "ggml-org SmolVLM 500M Instruct GGUF README", "url": "https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "runtime_format" ], "notes": "The README records Apache-2.0 licensing, base_model HuggingFaceTB/SmolVLM-500M-Instruct, and links to the llama.cpp SmolVLM GGUF support pull request." }, { "label": "HuggingFaceTB SmolVLM 500M Instruct API metadata", "url": "https://huggingface.co/api/models/HuggingFaceTB/SmolVLM-500M-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "base_total_params_b", "license", "pipeline" ], "notes": "The base repo API records Idefics3ForConditionalGeneration, image-text-to-text packaging, Apache-2.0 license, base_model HuggingFaceTB/SmolLM2-360M-Instruct plus google/siglip-base-patch16-512, region:us, and safetensors BF16 total 507482304 parameters. The base total includes vision/projector tensors that are stored in separate mmproj GGUF sidecars in the ggml-org package." }, { "label": "HuggingFaceTB SmolVLM 500M Instruct config", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct/raw/a7da5b986cb59b408707209984f360a5f4ad7e47/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Idefics3ForConditionalGeneration with a VLlama3ForCausalLM text decoder, BF16 dtype, untied embeddings, 32 text layers, hidden size 960, intermediate size 2560, 15 attention heads, 5 KV heads, 64 head dimension, 8192 max position embeddings, and a resident SigLIP-style vision config with 12 layers, hidden size 768, 12 heads, 512 image size, plus an Idefics3 projector." }, { "label": "ggml-org SmolVLM 500M Instruct GGUF linked sizes", "url": "https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF/tree/72e986006ef53e37cdd3f6d4241c90b0f01df376", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found SmolVLM-500M-Instruct-f16.gguf is 820422912 bytes, exactly matching API gguf.totalFileSize. Sibling files are SmolVLM-500M-Instruct-Q8_0.gguf at 436806912 bytes, mmproj-SmolVLM-500M-Instruct-f16.gguf at 199468800 bytes, and mmproj-SmolVLM-500M-Instruct-Q8_0.gguf at 108783360 bytes. The F16 mmproj sidecar header contains 98229504 logical parameters and 0.199455744 GB of tensor payload; it is not included in this selected-main-artifact profile." }, { "label": "ggml-org SmolVLM 500M Instruct F16 GGUF range-read tensor index", "url": "https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF/resolve/72e986006ef53e37cdd3f6d4241c90b0f01df376/SmolVLM-500M-Instruct-f16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 16MB range-read of the selected GGUF v3 header found 49 metadata entries and 291 tensors. The selected file is 0.820422912 GB, with tensor payloads starting at byte 1792512. Tensor spans total 0.818630400 GB across 409252800 logical elements: output.weight 0.094617600 GB, token_embd.weight 0.094617600 GB, blk.* tensors 0.629391360 GB, and output_norm.weight 0.000003840 GB. Metadata/tokenizer/header/file overhead accounts for 0.001792512 GB. Stored tensor bytes split into F16 0.818380800 GB and F32 0.000249600 GB. The header records general.architecture llama, general.license apache-2.0, llama.block_count 32, context_length 8192, embedding_length 960, feed_forward_length 2560, attention.head_count 15, attention.head_count_kv 5, attention key/value length 64, rope.freq_base 100000, vocab_size 49280, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, README, base model API/config, linked-object HEAD checks for all GGUF files, direct selected main GGUF header range read, and direct mmproj F16 sidecar header review." }, "notes": "Use this profile for the ggml-org API-selected F16 main GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--diffusiongemma-26b-a4b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/diffusiongemma-26B-A4B-it", "title": "Google DiffusionGemma 26B A4B IT BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 DiffusionGemma 26B A4B instruction-tuned repo.", "model_family": "diffusion-gemma-block-diffusion-moe", "architecture": { "canonical_architecture_id": "diffusion-gemma-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 51.647562456, "main_resident_weight_gb": 50.501973564, "auxiliary_resident_weight_gb": 1.145588892, "fixed_weight_gb": 4.826003004, "routed_expert_weight_gb": 0.35684352, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "Exact decoder tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because DiffusionGemma block diffusion is not ordinary one-output-token autoregressive decode.", "auxiliary_scope": "model.encoder tensors are resident for the multimodal encoder/cache path but are not enough to define ordinary token-by-token swept traffic.", "shared_expert_notes": "The config records top_k_experts 8 and 128 routed experts. The checkpoint also stores dense model.decoder.layers.*.mlp.* tensors outside model.decoder.layers.*.experts.*, so those always-on/shared tensors are included in fixed_weight_gb.", "notes": "Header-derived bytes are used. model.decoder tensors total 50.501973564 GB, model.encoder tensors total 1.145588892 GB, and no lm_head tensor is stored separately. Routed expert tensors total 45.67597056 GB and divide exactly into 128 uniform expert groups of 0.35684352 GB." }, "kv_adapter": { "kind": "unknown", "reason": "DiffusionGemma uses block diffusion over a 256-token canvas with a decoder that applies bidirectional attention over the generation canvas and then appends fully denoised canvases to cache. Bounds Engine v1 only models ordinary autoregressive per-output-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "A production profile needs a dedicated block-diffusion adapter with canvas length, denoising iteration count, sampler behavior, canvas self-attention traffic, cross-attention/context-cache traffic, and block append policy. Do not reuse the Gemma 4 autoregressive KV adapter for this repo." }, "notes": "This profile intentionally fails closed even though config and tensor headers are accessible, because the supported comparison math does not model DiffusionGemma's block-diffusion generation algorithm." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-block-diffusion", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The config dtype is bfloat16, and direct safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Google DiffusionGemma 26B A4B IT API metadata", "url": "https://huggingface.co/api/models/google/diffusiongemma-26B-A4B-it", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 0f28bc42f588fbd8f71e08102b1c3960298a1358, the API reports an Apache-2.0 image-text-to-text repo with diffusion_gemma tags, 1,424,776 downloads, region:us, and BF16 safetensors metadata." }, { "label": "Google DiffusionGemma 26B A4B IT config", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it/raw/0f28bc42f588fbd8f71e08102b1c3960298a1358/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "unsupported_reason" ], "notes": "The config records DiffusionGemmaForBlockDiffusion, model_type diffusion_gemma, bfloat16 dtype, canvas_length 256, 30 decoder layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 16 attention heads, 8 local KV heads, 2 global KV heads, 128 experts, top_k_experts 8, 262144 max position embeddings, and use_bidirectional_attention set for the vision/canvas path." }, { "label": "Google DiffusionGemma 26B A4B IT model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The model card describes a block-autoregressive architecture that iteratively denoises a 256-token canvas in parallel, then processes the completed canvas into the cache. That generation algorithm is not equivalent to ordinary one-token autoregressive decode." }, { "label": "Google DiffusionGemma 26B A4B IT safetensors index and shard headers", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it/raw/0f28bc42f588fbd8f71e08102b1c3960298a1358/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 51647562456 bytes across 11 shards. Range-read safetensors headers found 1047 BF16 tensors totaling 25.823781228 params / 51.647562456 GB. model.decoder tensors total 50.501973564 GB; model.encoder tensors total 1.145588892 GB. model.decoder.embed_tokens.weight is 1.476395008 GB, and there is no separate lm_head tensor. Non-expert decoder tensors total 4.826003004 GB. Routed expert tensors under model.decoder.layers.*.experts.down_proj and gate_up_proj total 45.67597056 GB and divide exactly into 128 uniform expert groups of 0.35684352 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Reviewed from the public model card, served config, HF API metadata, safetensors index, and direct shard header byte grouping. Marked unsupported because Bounds Engine v1 lacks a block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model block-diffusion generation over a denoised canvas, so ordinary autoregressive throughput would be misleading even though resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated DiffusionGemma block-diffusion adapter exists." }, { "id": "google--gemma-1-1-2b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-1.1-2b-it", "title": "Google Gemma 1.1 2B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 1.1 2B instruction-tuned repo.", "model_family": "gemma-dense", "base_model_proof": { "base_model": "unknown", "relation": "derived_package", "source": "Hugging Face API metadata and access checks only", "config_compatible": false, "notes": "The public API metadata does not expose a base_model field for this repo, and a google/gemma-1.1-2b base repo was not available through the API or HF CLI in this audit environment. Treat this as unverified lineage metadata, not an audited architecture comparison." }, "architecture": { "canonical_architecture_id": "gemma-1.1-2b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.506172416, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2506172416 BF16 safetensors parameters for this repo. KV geometry, context length, tied embedding behavior, and swept ordinary text-decode traffic are not audited because the gated config and safetensors headers are inaccessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw URL plus hf download access checks for config, README, generation config, safetensors index, and safetensors shards returned access denied in this audit environment.", "notes": "Do not infer Gemma layer count, KV heads, context length, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 1.1 2B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-1.1-2b-it", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit d750f5eceb83e978c09e2b3597c2a8784e381022, the API reports gated: manual, text-generation pipeline, Gemma license, GemmaForCausalLM/model_type gemma metadata, BF16 safetensors count 2506172416, 198323 downloads, endpoints_compatible, and region:us. The API exposes only a partial top-level config and tokenizer metadata, not layer count, KV heads, context length, tied embeddings, or tensor grouping." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-1.1-2b-it/raw/d750f5eceb83e978c09e2b3597c2a8784e381022/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw requests for config.json, model.safetensors.index.json, and safetensors HEAD/range access returned 401 GatedRepo responses. hf download google/gemma-1.1-2b-it for config.json, README.md, generation_config.json, model.safetensors.index.json, and both safetensors shards with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." }, { "label": "Base-repo availability check", "url": "https://huggingface.co/api/models/google/gemma-1.1-2b", "source_type": "manual_review", "supports": [ "base_model_proof", "unsupported_reason" ], "notes": "The API request for google/gemma-1.1-2b returned not found, and hf download google/gemma-1.1-2b for config and index files also returned not found for the configured CLI identity. This profile therefore does not infer architecture from an unavailable base repo." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from live HF API metadata, unauthenticated raw URL failures, authenticated HF CLI access checks, and base-repo availability checks." }, "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-2-2b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-2-2b-it", "title": "Google Gemma 2 2B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 2 2B instruction-tuned repo.", "model_family": "gemma2-dense", "base_model_proof": { "base_model": "google/gemma-2-2b", "relation": "finetune", "source": "Hugging Face API base_model metadata only", "config_compatible": false, "notes": "The API metadata records google/gemma-2-2b as the base model, but both the instruction-tuned repo and base repo configs are gated in this audit environment. Treat this as lineage metadata, not an audited architecture comparison." }, "architecture": { "canonical_architecture_id": "gemma-2-2b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.614341888, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2614341888 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo and base repo require approval, and hf download for config, README, generation config, and safetensors index returned access denied in this audit environment.", "notes": "Do not infer Gemma 2 layer count, KV heads, context length, sliding-window behavior, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 2 2B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-2b-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8, the API reports gated: manual, text-generation pipeline, Gemma license, base_model google/gemma-2-2b, BF16 safetensors count 2614341888, 445661 downloads, endpoints_compatible, and region:us." }, { "label": "Google Gemma 2 2B base API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-2b", "source_type": "model_card", "supports": [ "base_model_metadata", "unsupported_reason" ], "notes": "At commit c5ebcd40d208330abc697524c919956e692655cf, the base repo API also reports gated: manual and safetensors total 2614341888 parameters. The base repo API records F32 storage metadata, but the direct config and tensor index remain inaccessible here." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-2-2b-it/raw/299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "hf download google/gemma-2-2b-it for config.json, generation_config.json, model.safetensors.index.json, and README.md returned access denied because the repository requires approval. hf download google/gemma-2-2b for config.json, model.safetensors.index.json, and README.md returned the same access-denied result." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-2-2b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-2-2b", "title": "Google Gemma 2 2B F32", "summary": "Unsupported profile stub for the gated F32 Gemma 2 2B base repo.", "model_family": "gemma2-dense", "base_model_proof": { "base_model": "google/gemma-2-2b", "relation": "base", "source": "Hugging Face API metadata only", "config_compatible": false, "notes": "The API metadata identifies this as a Gemma2ForCausalLM base repo, but config and tensor headers are gated in this audit environment. Treat this as high-level API metadata, not an audited architecture comparison." }, "architecture": { "canonical_architecture_id": "gemma-2-2b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.614341888, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2614341888 F32 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and hf download for config, README, and safetensors index returned access denied in this audit environment.", "notes": "Do not infer Gemma 2 layer count, KV heads, context length, sliding-window behavior, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "F32 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 2 2B API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-2b", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit c5ebcd40d208330abc697524c919956e692655cf, the API reports gated: manual, text-generation pipeline, Gemma license, Gemma2ForCausalLM, F32 safetensors count 2614341888, 270278 downloads, endpoints_compatible, and region:us." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-2-2b/raw/c5ebcd40d208330abc697524c919956e692655cf/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config access returned HTTP 401 restricted access. hf download google/gemma-2-2b for config.json, README.md, and model.safetensors.index.json with the configured CLI token returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-2-9b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-2-9b-it", "title": "Google Gemma 2 9B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 2 9B IT repo.", "model_family": "gemma2-dense", "architecture": { "canonical_architecture_id": "gemma-2-9b-it", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 9.241705984, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 9241705984 BF16 safetensors parameters for this repo. KV geometry, context length, layer count, head layout, and tied-embedding layout are not audited because the gated config and model card are inaccessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is manually gated. Unauthenticated raw config requests return 401, and hf download with the configured CLI token returns access denied because the repository requires approval.", "notes": "Do not infer Gemma 2 layer count, KV heads, head dimension, context length, or tied embedding layout from the model name, base-model name, or scrape metadata. Replace this with an audited adapter only after direct config and tensor evidence are available." }, "notes": "This profile intentionally fails closed until the gated config and tensor layout can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and ordinary text-decode traffic cannot be verified.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 2 9B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-9b-it", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "license", "pipeline", "unsupported_reason" ], "notes": "At commit 11c9b309abf73637e4b6f9a3fa1e92e615547819, the API reports gated: manual, text-generation pipeline, Gemma license, region:us, base_model google/gemma-2-9b, BF16 safetensors count 9241705984, and 247024 downloads." }, { "label": "Gated config and model-card access check", "url": "https://huggingface.co/google/gemma-2-9b-it/raw/11c9b309abf73637e4b6f9a3fa1e92e615547819/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config and README requests returned 401 restricted access. `hf download google/gemma-2-9b-it config.json --revision 11c9b309abf73637e4b6f9a3fa1e92e615547819` with the configured CLI token returned access denied because the repository requires approval. The base repo `google/gemma-2-9b` returned the same approval requirement for config access." } ], "unsupported_reason": "Gated config and model card are not accessible in this audit environment, so KV geometry, context length, and ordinary text-decode weight traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and tensor evidence are available." }, { "id": "google--gemma-2b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-2b", "title": "Google Gemma 2B BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 2B base repo.", "model_family": "gemma-dense", "base_model_proof": { "base_model": "google/gemma-2b", "relation": "base", "source": "Hugging Face API metadata, accessible model card, and gated access checks", "config_compatible": false, "notes": "The repo is itself the Gemma 2B base model. The model card metadata now points to google/gemma-2-2b as a newer version, but this profile intentionally records the legacy google/gemma-2b repo without substituting the newer architecture." }, "architecture": { "canonical_architecture_id": "gemma-2b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense", "total_params_b": 2.506172416, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2506172416 BF16 safetensors parameters for this repo. The model card records an 8192-token trained context length. KV geometry, tied embedding behavior, and swept ordinary text-decode traffic are not audited because the gated config, safetensors index, safetensors shards, and GGUF object are inaccessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw URL plus authenticated HF CLI access checks for config, generation config, safetensors index, safetensors shard, and the GGUF object returned access denied in this audit environment.", "notes": "Do not infer Gemma layer count, KV heads, head dimension, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata. The repo has a gemma-2b.gguf sibling, but that object is gated from the configured HF token and is not modeled." }, "evidence": [ { "label": "Google Gemma 2B API metadata", "url": "https://huggingface.co/api/models/google/gemma-2b", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 9cf48e52b224239de00d483ec8eb84fb8d0f3a3a, the API reports gated: manual, text-generation pipeline, Gemma license, GemmaForCausalLM/model_type gemma metadata, BF16 safetensors count 2506172416, 117744 downloads, endpoints_compatible, and region:us. The API exposes only partial top-level config/tokenizer metadata, not layer count, KV heads, head dimension, tensor grouping, or tied embedding behavior." }, { "label": "Google Gemma 2B model card", "url": "https://huggingface.co/google/gemma-2b", "source_type": "model_card", "supports": [ "max_context_tokens", "repo", "new_version_metadata" ], "notes": "The authenticated HF CLI could download README.md. The card identifies this as the 2B base version of Gemma, records that the models were trained on an 8192-token context length, and has cardData.new_version google/gemma-2-2b. It does not expose the gated config or tensor layout needed for production bounds." }, { "label": "Gated config and tensor access checks", "url": "https://huggingface.co/google/gemma-2b/raw/9cf48e52b224239de00d483ec8eb84fb8d0f3a3a/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw requests for config.json, model.safetensors.index.json, and the pinned gemma-2b.gguf HEAD returned HTTP 401 GatedRepo responses. Authenticated hf download with the configured osolmaz identity returned Access denied for config.json, generation_config.json, model.safetensors.index.json, model-00001-of-00002.safetensors, and gemma-2b.gguf. README.md was the only checked artifact that downloaded." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from live HF API metadata, authenticated HF CLI access checks, unauthenticated raw URL failures, and accessible model-card context metadata." }, "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and tensor header evidence is available." }, { "id": "google--gemma-3-12b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-12b-it", "title": "Google Gemma 3 12B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 12B instruction-tuned repo.", "model_family": "gemma3-dense", "architecture": { "canonical_architecture_id": "gemma-3-12b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 12.18732504, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 12187325040 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, generation config, safetensors index, safetensors bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Gemma 3 layer count, KV heads, context length, sliding-window behavior, vision/text split, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 12B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-12b-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 96b6f1eccf38110c56df3a15bffe176da04bfd80, the API reports gated: manual, image-text-to-text pipeline, Gemma license, base_model google/gemma-3-12b-pt, BF16 safetensors count 12187325040, and current downloads 1705245." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-3-12b-it/raw/96b6f1eccf38110c56df3a15bffe176da04bfd80/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config, README, generation config, safetensors index, and ranged safetensors requests returned 401 access-denied responses. hf download config.json with the configured CLI token returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, vision/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-3-1b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-1b-it", "title": "Google Gemma 3 1B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 1B instruction-tuned repo.", "model_family": "gemma3-dense", "architecture": { "canonical_architecture_id": "gemma-3-1b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 0.999885952, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 999885952 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, safetensors index, safetensors range reads, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Gemma 3 layer count, KV heads, context length, sliding-window behavior, or RoPE settings from the model name or scraped metadata. Replace this with an audited adapter only after direct config evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 1B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-1b-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "The API reports text-generation pipeline, Gemma license tags, base_model google/gemma-3-1b-pt, BF16 safetensors count 999885952, commit dcc83ea841ab6100d6b47a070329e1ba4cf78752, and lastModified 2025-04-04T13:12:40Z." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-3-1b-it/raw/main/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config and README requests returned 401. Raw safetensors index returned 401. A ranged request for model.safetensors returned a 401 access-denied message rather than safetensors bytes. hf download config.json with the configured osolmaz CLI token returned access denied because the repository requires approval. The base repo google/gemma-3-1b-pt config was also access denied." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-3-270m-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-270m-it", "title": "Google Gemma 3 270M IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 270M instruction-tuned repo.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-270m", "relation": "finetune", "source": "Hugging Face API metadata", "config_compatible": false, "notes": "The API metadata identifies google/gemma-3-270m as the base model, but direct config comparison is not possible in this environment because both repos require manual gated approval." }, "architecture": { "canonical_architecture_id": "gemma-3-270m", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 0.268098176, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 268098176 BF16 safetensors parameters for this repo. KV geometry is not audited because the gated config is not accessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is manually gated, unauthenticated raw config access returns 401, and hf download with the configured CLI token returns access denied because the repository requires approval.", "notes": "Do not infer Gemma layer count, KV heads, attention pattern, or context length from the model name or scrape metadata. Replace this with an audited adapter only after direct config access is available." }, "notes": "This profile intentionally fails closed until the gated config can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 270M IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-270m-it", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit ac82b4e820549b854eebf28ce6dedaf9fdfa17b3, the API reports gated: manual, text-generation pipeline, Gemma license, region:us, 1035190 downloads, base_model google/gemma-3-270m, and BF16 safetensors count 268098176." }, { "label": "Gated raw config access check", "url": "https://huggingface.co/google/gemma-3-270m-it/raw/ac82b4e820549b854eebf28ce6dedaf9fdfa17b3/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config access returned 401 Unauthorized with a restricted-repo message." }, { "label": "Gated HF CLI config access check", "url": "https://huggingface.co/google/gemma-3-270m-it", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "hf auth whoami reports the CLI is logged in as osolmaz, but hf download of config.json at the pinned commit returned: Access denied. This repository requires approval." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata and direct access checks. Production bounds intentionally remain disabled because the served config and KV geometry are not accessible." }, "unsupported_reason": "Gated config is not accessible in this audit environment, so KV geometry and max context cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config evidence is available." }, { "id": "google--gemma-3-270m", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-270m", "title": "Google Gemma 3 270M BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 270M repo.", "model_family": "gemma3-dense", "architecture": { "canonical_architecture_id": "gemma-3-270m", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 0.268098176, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 268098176 BF16 safetensors parameters for this repo. KV geometry is not audited because the gated config is not accessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is manually gated, and hf download config.json with the configured CLI token returned access denied because the repository requires approval.", "notes": "Do not infer Gemma layer count, KV heads, attention pattern, or context length from the model name or the scrape metadata. Replace this with an audited adapter only after direct config access is available." }, "notes": "This profile intentionally fails closed until the gated config can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 270M API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-270m", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "The API reports gated: manual, text-generation pipeline, Gemma license, and BF16 safetensors count 268098176." }, { "label": "Gated config access check", "url": "https://huggingface.co/google/gemma-3-270m/raw/main/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config returned 401, and hf download with the configured CLI token returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config is not accessible in this audit environment, so KV geometry and max context cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config evidence is available." }, { "id": "google--gemma-3-27b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-27b-it", "title": "Google Gemma 3 27B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 27B instruction-tuned repo.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-27b-pt", "relation": "finetune", "source": "Hugging Face API cardData and base_model tags", "config_compatible": false, "notes": "The public API records google/gemma-3-27b-pt as the base model, but raw config and tensor headers are gated for both the instruction-tuned and base repos in this audit environment." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 27.43240664, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 27432406640 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, generation config, safetensors index, safetensors bytes, base-repo config/index, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Gemma 3 layer count, KV heads, context length, sliding-window behavior, vision/text split, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 27B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-27b-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 005ad3404e59d6023443cb575daa05336842228a, the API reports gated: manual, image-text-to-text pipeline, Gemma license, base_model google/gemma-3-27b-pt, BF16 safetensors count 27432406640, current downloads 962269, endpoints_compatible, and region:us." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-3-27b-it/raw/005ad3404e59d6023443cb575daa05336842228a/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Instruction-tuned raw config, README, generation_config, and safetensors index returned 401 access-denied responses. The base repo google/gemma-3-27b-pt raw config, README, and safetensors index also returned 401. hf download config.json with the configured osolmaz CLI token returned access denied because the repository requires approval." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Reviewed from current public HF API metadata plus unauthenticated and configured-token access checks. Marked unsupported because direct config and tensor-header evidence is unavailable." }, "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, vision/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-3-4b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3-4b-it", "title": "Google Gemma 3 4B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3 4B instruction-tuned repo.", "model_family": "gemma3-dense", "architecture": { "canonical_architecture_id": "gemma-3-4b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 4.300079472, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 4300079472 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, generation config, safetensors index, safetensors bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Gemma 3 layer count, KV heads, context length, sliding-window behavior, vision/text split, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3 4B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-4b-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 093f9f388b31de276ce2de164bdc2081324b9767, the API reports gated: manual, image-text-to-text pipeline, Gemma license, base_model google/gemma-3-4b-pt, BF16 safetensors count 4300079472, and current downloads 1795040." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-3-4b-it/raw/093f9f388b31de276ce2de164bdc2081324b9767/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config, README, generation config, safetensors index, and ranged safetensors requests returned 401 GatedRepo access-denied responses. hf download config.json with the configured CLI token returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, vision/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-3n-e2b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-3n-E2B-it", "title": "Google Gemma 3n E2B IT BF16", "summary": "Unsupported profile stub for the gated BF16 Gemma 3n E2B instruction-tuned multimodal repo.", "model_family": "gemma3n-multimodal", "architecture": { "canonical_architecture_id": "gemma-3n-e2b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 5.439438272, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 5439438272 BF16 safetensors parameters for this repo. KV geometry, text/multimodal tensor split, and swept ordinary text traffic are not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, safetensors index, safetensors shard bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Gemma 3n layer count, local/global attention behavior, KV heads, context length, multimodal encoders, per-layer embeddings, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google Gemma 3n E2B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3n-E2B-it", "source_type": "model_card", "supports": [ "repo", "base_model_metadata", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 5e092ebca197cdcd8d8b195040accf22693501bc, the API reports gated: manual, Transformers image-text-to-text pipeline, Gemma license, base_model google/gemma-3n-E4B-it with finetune relation, BF16 safetensors count 5439438272, current downloads 409872, endpoints_compatible, and region:us." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/gemma-3n-E2B-it/raw/5e092ebca197cdcd8d8b195040accf22693501bc/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "hf download config.json with the configured CLI token for osolmaz returned access denied because the repository requires approval. Raw pinned config and ranged pinned safetensors shard requests returned 401 GatedRepo responses, so exact architecture, context, KV geometry, and tensor grouping cannot be verified." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context, multimodal/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--gemma-4-12b-it-qat-q4-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-12B-it-qat-q4_0-gguf", "title": "Google Gemma 4 12B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for Google's official Q4_0 GGUF artifact of Gemma 4 12B IT QAT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 12B text geometry as the Google QAT unquantized config. The Google official GGUF repo ships only GGUF files, so the immutable Google QAT unquantized config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 6.975877728, "swept_weight_gb": 6.960055808, "auxiliary_resident_weight_gb": 0.01582192, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-12b-it-qat-q4_0.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as model tensors; mmproj-gemma-4-12b-it-qat-q4_0.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_0 GGUF file selected by HF API gguf.totalFileSize. Header tensor payloads total 6.960054464 GB, with 0.000001344 GB of tensor-alignment padding, while the linked file size is 6.975877728 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5858463378125703, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is Google's official Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-12b-it-qat-q4_0.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Google Gemma 4 12B IT QAT Q4_0 GGUF HF API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-q4_0-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit f6e7774e6148da3b7f201e42ba37cf084c1db35f records base_model google/gemma-4-12B-it-qat-q4_0-unquantized, Apache-2.0 license, any-to-any pipeline, region:us, 564056 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 6975877728." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-12B-it-qat-q4_0-unquantized, QAT release packaging, and describes Q4_0 GGUF artifacts as ready-to-deploy formats for broad ecosystem compatibility." }, { "label": "Google Gemma 4 12B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "Google Gemma 4 12B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf/tree/f6e7774e6148da3b7f201e42ba37cf084c1db35f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-12b-it-qat-q4_0.gguf is 6975877728 bytes, exactly matching API gguf.totalFileSize. The sidecar mmproj-gemma-4-12b-it-qat-q4_0.gguf is 175115264 bytes and is not the selected main text artifact." }, { "label": "Google Gemma 4 12B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf/resolve/f6e7774e6148da3b7f201e42ba37cf084c1db35f/gemma-4-12b-it-qat-q4_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 44 metadata entries and 667 tensors. The selected file is 6.975877728 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15821920. Tensor payloads total 6.960054464 GB with 0.000001344 GB of tensor-alignment padding, for swept tensor spans of 6.960055808 GB across 11907350576 logical elements: token_embd.weight 0.8257536 GB, blk.* tensors 6.13428448 GB, output_norm.weight 0.00001536 GB, and rope_freqs.weight 0.000001024 GB. Tensor payloads split into Q4_0 6.13122048 GB, Q6_K 0.8257536 GB, and F32 0.003080384 GB. Metadata/tokenizer/header/file overhead accounts for 0.01582192 GB. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for Google's official main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--gemma-4-12b-it-qat-q4-0-unquantized", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-12B-it-qat-q4_0-unquantized", "title": "Google Gemma 4 12B IT QAT Q4_0 Unquantized BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 unquantized QAT checkpoint of Gemma 4 12B unified instruction-tuned.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "finetune", "source": "Hugging Face model card base_model metadata, served QAT config, base instruction config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The repo records google/gemma-4-12B-it as its base model and stores the unquantized BF16 QAT checkpoint used by downstream Q4_0 packages. Manual comparison found the memory-relevant text decode fields match the base instruction model: Gemma4UnifiedForConditionalGeneration, BF16 dtype, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, tied embeddings, and 262144 context. The QAT config changes some multimodal/default metadata, so this profile uses the served QAT config and direct safetensors header as authoritative." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.959730224, "swept_params_b": 11.90735032, "auxiliary_resident_params_b": 0.052379904, "resident_weight_gb": 23.919460448, "swept_weight_gb": 23.81470064, "auxiliary_resident_weight_gb": 0.104759808, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model safetensors headers", "auxiliary_scope": "model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for multimodal inputs but not swept for each generated text token", "notes": "This unquantized QAT checkpoint stores only BF16 tensors. Gemma 4 12B Unified is encoder-free, so the multimodal auxiliary footprint is only the lightweight projection/embedding tensors outside model.language_model. The config records tie_word_embeddings true and the header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-qat-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this unquantized BF16 QAT repo. Downstream Q4_0 and W4A16 packages have separate profiles.", "notes": "The top-level config records dtype bfloat16, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 12B IT QAT unquantized model card and API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-q4_0-unquantized", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The HF CLI/API response records repo SHA 58540658b6c08edab2ddc1fbde7f28cc9987ced3, Apache-2.0 licensing, any-to-any pipeline, base_model google/gemma-4-12B-it, region:us, 218747 downloads, and safetensors logical parameters BF16: 11959730224, total: 11959730224. The model card metadata identifies this as the unquantized QAT checkpoint behind Q4_0 packages." }, { "label": "Google Gemma 4 12B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, num_global_key_value_heads 1, global_head_dim 512, 8 sliding KV heads, 256 sliding head dimension, 262144 max position embeddings, and lightweight audio/vision projection configs." }, { "label": "Google Gemma 4 12B IT base instruction config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the ordinary text decode fields match the QAT unquantized repo: architecture, model_type, dtype, tie_word_embeddings, layer count, layer types, attention geometry, sliding-window length, and max context. The QAT config changes some multimodal/default metadata, so this profile does not copy the base config wholesale." }, { "label": "Google Gemma 4 12B IT QAT unquantized safetensors header", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/resolve/58540658b6c08edab2ddc1fbde7f28cc9987ced3/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The file Content-Length is 23919549408 bytes, with an 88952-byte header and 23919460448 tensor bytes across 677 tensors. Stored tensors sum to 11.959730224B BF16 parameters / 23.919460448 GB. model.language_model tensors sum to 11.907350320B BF16 parameters / 23.814700640 GB and include model.language_model.embed_tokens.weight with shape [262144, 3840]. The header has no separate lm_head.weight, so the embedding table remains swept as the tied output projection. Resident-only tensors outside model.language_model sum to 0.052379904B BF16 parameters / 0.104759808 GB: model.embed_audio, model.embed_vision, and model.vision_embedder. Full-attention layer headers have no v_proj tensors, matching attention_k_eq_v true." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, served QAT config, base instruction config comparison, model card metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds on the unquantized BF16 QAT checkpoint. It deliberately separates resident multimodal projection weights from per-token swept language weights." }, { "id": "google--gemma-4-12b-it-qat-w4a16-ct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-12B-it-qat-w4a16-ct", "title": "Google Gemma 4 12B Unified IT QAT W4A16 CT", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors W4A16 QAT Gemma 4 12B unified instruction-tuned repo.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors repo records google/gemma-4-12B-it-qat-q4_0-unquantized as its base model, and that unquantized QAT repo records google/gemma-4-12B-it as its base. Manual comparison found matching text and audio decode geometry between the compressed and unquantized QAT configs. The compressed repo adds quantization_config plus explicit top-level and vision_config defaults that do not change the text-decode fields used by this profile." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 13.3069872, "swept_params_b": 12.247974336, "auxiliary_resident_params_b": 1.059012864, "resident_weight_gb": 10.264057056, "swept_weight_gb": 8.146031328, "auxiliary_resident_weight_gb": 2.118025728, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens input table, model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for the package but not full-matrix swept for each generated text token", "notes": "Compressed-tensors stores packed I32 int4 weights, BF16 scale and ignored-module tensors, and tiny I64 weight_shape side tensors. The config records tie_word_embeddings true, but this compressed artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection and treats model.language_model.embed_tokens.weight as a resident input table for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Quantizing weights does not change the BF16 KV cache assumption because the config has kv_cache_scheme null." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models ordinary text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7713283932519301, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scale tensors, ignored BF16 modules, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, bfloat16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records the exact resident stored-byte average across logical safetensors parameters; resident and swept traffic use the exact header byte fields above." }, "evidence": [ { "label": "Google Gemma 4 12B IT QAT W4A16 CT model card and API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-w4a16-ct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The HF CLI/API response records repo SHA 9c79b5e652ae36f02bb07d3ca29124a9d1b009bd, Apache-2.0 licensing, any-to-any pipeline, base_model google/gemma-4-12B-it-qat-q4_0-unquantized, compressed-tensors tags, and safetensors logical parameters I64: 656, I32: 10899947520, BF16: 2407039024, total: 13306987200. The model card describes the W4A16 compressed-tensors artifacts as QAT checkpoints serialized for native optimized inference with vLLM." }, { "label": "Google Gemma 4 12B IT QAT W4A16 CT config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-w4a16-ct/raw/9c79b5e652ae36f02bb07d3ca29124a9d1b009bd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, compressed-tensors pack-quantized int4 weights, group_size 32, symmetric group strategy, quantization_status compressed, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, num_global_key_value_heads 1, global_head_dim 512, 8 sliding KV heads, 256 sliding head dimension, 262144 max position embeddings, and lightweight audio/vision projection configs. The explicit quantization ignore list has 17 entries including lm_head, audio/embed/vision projection tensors, and vision embedder norm/dense tensors." }, { "label": "Google Gemma 4 12B IT QAT W4A16 CT quantization recipe", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-w4a16-ct/raw/9c79b5e652ae36f02bb07d3ca29124a9d1b009bd/recipe.yaml", "source_type": "config", "supports": [ "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The recipe records QuantizationModifier targets Linear with int num_bits 4, group_size 32, symmetric group strategy, memoryless_minmax observer, and ignore patterns for lm_head, embedding, vision, audio, and router modules." }, { "label": "Google Gemma 4 12B IT QAT unquantized base config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the audited text_config and audio_config geometry fields match the compressed-tensors repo. The compressed repo adds quantization_config and explicit top-level hidden_size/intermediate_size plus vision_config model_patch_size/num_soft_tokens defaults; these do not change the ordinary text-decode architecture fields." }, { "label": "Google Gemma 4 12B IT QAT W4A16 CT safetensors header", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-w4a16-ct/resolve/9c79b5e652ae36f02bb07d3ca29124a9d1b009bd/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The file Content-Length is 10264229896 bytes, with a 172832-byte header and 10264057056 tensor bytes across 1334 tensors. Stored tensors sum to 13.3069872B logical parameters / 10.264057056 GB: 328 I32 packed tensors carrying 10.89994752B logical int4 weights / 5.44997376 GB, 678 BF16 tensors carrying 2.407039024B parameters / 4.814078048 GB, and 328 I64 shape tensors carrying 656 values / 0.000005248 GB. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens.weight plus lm_head.weight, sum to 12.247974336B logical parameters / 8.146031328 GB. Resident-only tensors, defined as model.language_model.embed_tokens.weight plus model.embed_audio, model.embed_vision, and model.vision_embedder tensors, sum to 1.059012864B parameters / 2.118025728 GB. Full-attention layer headers have no v_proj tensors, matching attention_k_eq_v true." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF CLI/API metadata, the served config, quantization recipe, unquantized QAT base config comparison, model card, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident input embedding and multimodal projection tensors from per-token swept language/logit weights." }, { "id": "google--gemma-4-12b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-12B-it", "title": "Google Gemma 4 12B Unified IT BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 12B unified instruction-tuned repo.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B", "relation": "finetune", "source": "Hugging Face model card base_model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text, audio, vision, dtype, context, and attention fields between the instruction-tuned and base configs. The instruction-tuned config adds a top-level eos_token_id list, which does not change the memory-side architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.959730224, "swept_params_b": 11.90735032, "auxiliary_resident_params_b": 0.052379904, "resident_weight_gb": 23.919460448, "swept_weight_gb": 23.81470064, "auxiliary_resident_weight_gb": 0.104759808, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for multimodal inputs but not swept for each generated text token", "notes": "Gemma 4 12B Unified is encoder-free, so the multimodal auxiliary footprint is only the lightweight projection/embedding tensors outside model.language_model. The config records tie_word_embeddings true and the header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level config records dtype bfloat16, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 12B IT model card", "url": "https://huggingface.co/google/gemma-4-12B-it", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "layers", "max_context_tokens", "unified_multimodal" ], "notes": "The model metadata identifies google/gemma-4-12B as the base model, Apache-2.0 licensing, and any-to-any multimodal packaging. The card states the 12B Unified model has 11.95B parameters, 48 layers, 256K context, and an encoder-free architecture where raw image patches and audio waveforms are projected into the LLM embedding space through lightweight linear layers." }, { "label": "Google Gemma 4 12B IT config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, tie_word_embeddings true, dtype bfloat16, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, num_global_key_value_heads 1, global_head_dim 512, 8 sliding KV heads, 256 sliding head dimension, 262144 max position embeddings, and lightweight audio/vision projection configs." }, { "label": "Google Gemma 4 12B base config", "url": "https://huggingface.co/google/gemma-4-12B/raw/56820d7d8cbe8e47975a53325439ed272e91cff2/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text, multimodal, serving, and context architecture fields match the instruction-tuned repo config." }, { "label": "Google Gemma 4 12B IT Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "base_model_proof", "total_params_b" ], "notes": "The HF CLI model info response records repo SHA 5926caa4ec0cac5cbfadaf4077420520de1d5205, safetensors parameters BF16: 11959730224, total: 11959730224, any-to-any pipeline, Apache-2.0 licensing, and base_model google/gemma-4-12B." }, { "label": "Google Gemma 4 12B IT safetensors header", "url": "https://huggingface.co/google/gemma-4-12B-it/resolve/5926caa4ec0cac5cbfadaf4077420520de1d5205/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. The file Content-Length is 23919549408 bytes, with an 88952-byte header and 23919460448 tensor bytes. Stored tensors sum to 11959730224 BF16 parameters / 23.919460448 GB. model.language_model tensors sum to 11907350320 BF16 parameters / 23.81470064 GB and include model.language_model.embed_tokens.weight with shape [262144, 3840]. The header has no separate lm_head.weight, so the embedding table remains swept as the tied output projection. Resident-only tensors outside model.language_model sum to 52379904 BF16 parameters / 0.104759808 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF CLI model info, served config, base config comparison, model card, direct safetensors header byte grouping, and local scrape row." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal projection weights from per-token swept language weights." }, { "id": "google--gemma-4-12b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-12B", "title": "Google Gemma 4 12B Unified BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 12B unified base repo.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B", "relation": "base", "source": "Hugging Face model card/API metadata, served config, instruction-tuned config comparison, and direct safetensors header range read", "config_compatible": true, "notes": "This is the base Gemma 4 12B Unified repo. Manual comparison against google/gemma-4-12B-it found matching profile-relevant architecture fields across the top-level architecture, dtype, tie_word_embeddings, text_config, audio_config, and vision_config. The instruction-tuned repo adds chat-template packaging and a top-level eos_token_id list, which does not change the memory-side fields used here." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.959730224, "swept_params_b": 11.90735032, "auxiliary_resident_params_b": 0.052379904, "resident_weight_gb": 23.919460448, "swept_weight_gb": 23.81470064, "auxiliary_resident_weight_gb": 0.104759808, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for multimodal inputs but not swept for each generated text token", "notes": "Gemma 4 12B Unified is encoder-free, so the multimodal auxiliary footprint is only the lightweight projection/embedding tensors outside model.language_model. The config records tie_word_embeddings true and the header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level config records dtype bfloat16, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 12B model card", "url": "https://huggingface.co/google/gemma-4-12B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "layers", "max_context_tokens", "unified_multimodal" ], "notes": "The card identifies this as the Gemma 4 12B Unified base model, with Apache-2.0 licensing, multimodal text/image/audio/video input packaging, 11.95B parameters, 48 layers, 256K context, encoder-free multimodal projection into the LLM embedding space, and hybrid local/global attention with unified K/V on global layers." }, { "label": "Google Gemma 4 12B config", "url": "https://huggingface.co/google/gemma-4-12B/raw/56820d7d8cbe8e47975a53325439ed272e91cff2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, tie_word_embeddings true, dtype bfloat16, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, num_global_key_value_heads 1, global_head_dim 512, 8 sliding KV heads, 256 sliding head dimension, 262144 max position embeddings, and lightweight audio/vision projection configs." }, { "label": "Google Gemma 4 12B IT config comparison", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "derived_calculation", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual JSON comparison found no differences in the profile-relevant architecture fields: architectures, dtype, model_type, tie_word_embeddings, text_config, audio_config, and vision_config. This matches the already audited Gemma 4 12B IT architecture." }, { "label": "Google Gemma 4 12B Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "weight_format", "total_params_b" ], "notes": "The HF CLI/API response at repo SHA 56820d7d8cbe8e47975a53325439ed272e91cff2 records a public non-gated Transformers any-to-any repo with safetensors, gemma4_unified, image-text-to-text, Apache-2.0 license, endpoints_compatible, region:us, 394761 downloads, and BF16 safetensors total 11959730224 parameters." }, { "label": "Google Gemma 4 12B safetensors header", "url": "https://huggingface.co/google/gemma-4-12B/resolve/56820d7d8cbe8e47975a53325439ed272e91cff2/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. The file Content-Length is 23919549408 bytes, with an 88952-byte header and 23919460448 tensor bytes. The header records 677 BF16 tensors totaling 11959730224 parameters / 23.919460448 GB. model.language_model tensors sum to 11907350320 BF16 parameters / 23.81470064 GB and include model.language_model.embed_tokens.weight with shape [262144, 3840]. The header has no separate lm_head.weight, so the embedding table remains swept as the tied output projection. Resident-only tensors outside model.language_model sum to 52379904 BF16 parameters / 0.104759808 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, model card, instruction-tuned config comparison, direct safetensors header byte grouping, and local scrape row." }, "notes": "This profile is for ordinary text decode bounds on the base Gemma 4 12B Unified checkpoint. It deliberately separates resident multimodal projection weights from per-token swept language weights." }, { "id": "google--gemma-4-26b-a4b-it-assistant", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-4-26B-A4B-it-assistant", "title": "Google Gemma 4 26B A4B IT Assistant Drafter BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 Gemma 4 26B A4B IT MTP assistant drafter repo.", "model_family": "gemma4-assistant-speculative-drafter", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "derived_package", "source": "Model card speculative-decoding instructions and generation_config assistant metadata", "config_compatible": false, "notes": "The model card identifies this repo as the Multi-Token Prediction drafter for google/gemma-4-26B-A4B-it. It is loaded as assistant_model alongside the target 26B A4B model in Transformers speculative decoding, not as a replacement target model." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b-assistant-drafter", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.419711236, "swept_params_b": 0.419711236, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.839422472, "swept_weight_gb": 0.839422472, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "Exact BF16 tensor bytes are recorded, but Bounds Engine v1 does not use them for production throughput because this repo is a speculative drafter paired with a target model.", "auxiliary_scope": "No separate auxiliary resident tensors were identified in the single safetensors file; the repo stores model.embed_tokens, model.layers, model.norm, pre_projection, and post_projection tensors.", "notes": "The single safetensors file stores 48 BF16 tensors totaling 419711236 parameters / 0.839422472 GB: model.embed_tokens 0.536870912 GB, model.layers 0.285248008 GB, model.norm 0.000002048 GB, pre_projection.weight 0.011534336 GB, and post_projection.weight 0.005767168 GB." }, "kv_adapter": { "kind": "unknown", "reason": "Gemma4AssistantForCausalLM is an MTP speculative drafter. Production throughput depends on the paired target model, draft length, acceptance rate, verification schedule, and extra target-side validation work. Bounds Engine v1 only models single-model ordinary autoregressive text-token decode.", "notes": "The config records four assistant text layers with three sliding-attention layers and one full-attention layer, but using those fields as ordinary target-model decode would be misleading because generation_config marks is_assistant true and num_assistant_tokens 6." }, "notes": "This profile intentionally fails closed even though exact weights and assistant-layer geometry are accessible, because comparison tok/s should not treat this drafter as a standalone target model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-gemma4-mtp-assistant-drafter", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo is used as assistant_model in speculative decoding with google/gemma-4-26B-A4B-it. A production profile needs a paired target+assistant speculative-decoding adapter, not a standalone ordinary text-decode adapter." }, "evidence": [ { "label": "Google Gemma 4 26B A4B IT Assistant API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-assistant", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 44033eb52554fc5339b383c81309119c684e4fb0, the API reports an Apache-2.0 any-to-any repo with transformers, safetensors, gemma4_assistant, text-generation, endpoints_compatible, region:us, 318820 downloads, and BF16 safetensors metadata for 419711236 parameters." }, { "label": "Google Gemma 4 26B A4B IT Assistant config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/raw/44033eb52554fc5339b383c81309119c684e4fb0/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "assistant_geometry", "unsupported_reason" ], "notes": "The config records Gemma4AssistantForCausalLM, model_type gemma4_assistant, BF16 dtype, backbone_hidden_size 2816, 2048 centroids, centroid_intermediate_top_k 32, use_ordered_embeddings false, tie_word_embeddings true, and a four-layer text_config with three sliding-attention layers, one full-attention layer, hidden_size 1024, intermediate_size 8192, 16 attention heads, 8 local KV heads, 2 global KV heads, attention_k_eq_v true, 1024-token sliding window, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B IT Assistant generation config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/raw/44033eb52554fc5339b383c81309119c684e4fb0/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config explicitly records is_assistant true, num_assistant_tokens 6, and a constant assistant-token schedule." }, { "label": "Google Gemma 4 26B A4B IT Assistant model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm", "base_model_proof" ], "notes": "The model card states that this checkpoint contains Multi-Token Prediction drafters for Gemma 4 models. Its examples load google/gemma-4-26B-A4B-it as TARGET_MODEL_ID and this repo as ASSISTANT_MODEL_ID, then pass assistant_model=assistant_model to target_model.generate." }, { "label": "Google Gemma 4 26B A4B IT Assistant safetensors header", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/resolve/44033eb52554fc5339b383c81309119c684e4fb0/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "resident_weight_gb", "weight_format" ], "notes": "A range-read of the single safetensors header found a 5360-byte header and 48 BF16 tensors totaling 419711236 parameters / 0.839422472 GB. Tensor groups are model.embed_tokens 0.536870912 GB, model.layers 0.285248008 GB, model.norm 0.000002048 GB, pre_projection.weight 0.011534336 GB, and post_projection.weight 0.005767168 GB. The linked file size is 839427840 bytes, so non-tensor safetensors overhead is 5368 bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from live HF API metadata, model card, served config, generation config, and direct safetensors header byte grouping. Marked unsupported because Bounds Engine v1 lacks a paired target+assistant speculative decoding adapter." }, "unsupported_reason": "Gemma 4 assistant checkpoints are speculative drafters used with a separate target model. Bounds Engine v1 does not model paired target+assistant speculative decoding, acceptance rates, or target verification work, so standalone tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports paired speculative-decoding profiles." }, { "id": "google--gemma-4-26b-a4b-it-qat-q4-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf", "title": "Google Gemma 4 26B A4B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for Google's official Q4_0 GGUF artifact of Gemma 4 26B A4B IT QAT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, non-QAT Google config comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-26B-A4B-it-qat-q4_0-unquantized. Manual comparison found the QAT unquantized config and the already audited google/gemma-4-26B-A4B-it config have matching checked text, vision, context, tied-embedding, expert, and attention geometry fields; the QAT config only updates Transformers metadata." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 14.43936144, "main_resident_weight_gb": 14.423538808, "auxiliary_resident_weight_gb": 0.015822632, "fixed_weight_gb": 1.577172088, "routed_expert_weight_gb": 0.10036224, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for gemma-4-26B_q4_0-it.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors; gemma-4-26B-it-mmproj.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "shared_expert_notes": "The model card records 8 active / 128 total experts and 1 shared expert. The GGUF header stores dense blk.*.ffn_down/gate/up tensors and tiny ffn_down_exps.scale tensors outside the routed Q4_0 expert weight span, so those shared/always-on bytes are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects gemma-4-26B_q4_0-it.gguf. A GGUF v3 range-read found 658 tensors and 46 metadata entries. Tensor spans total 14.423538808 GB, while the linked file is 14.439361440 GB. Routed expert Q4_0 weight tensors total 12.846366720 GB and divide exactly by 128 expert indexes. Non-expert and always-on tensor spans total 1.577172088 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The config layer_types and GGUF head-count array show five full-attention layers. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile targets Google's selected main Q4_0 QAT text GGUF artifact. Multimodal projector sidecar bytes and execution are separate workload concerns and are not included unless explicitly loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5722379485550018, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is Google's official Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-26B_q4_0-it.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Google Gemma 4 26B A4B IT QAT Q4_0 GGUF HF API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-qat-q4_0-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit dfc00409adc70be497fee9c90bfe76b3ee130f2e records base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, image-text-to-text pipeline, region:us, 338211 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 25233142046, and gguf.totalFileSize 14439361440." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/raw/dfc00409adc70be497fee9c90bfe76b3ee130f2e/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, QAT release packaging, Q4_0 GGUF ready-to-deploy artifacts, and Gemma 4 medium-model 256K context. The Gemma 4 architecture table and sibling audited profiles identify the 26B A4B model as 25.2B total, 3.8B active, 30 layers, 1024-token sliding window, and 8 active / 128 total experts plus 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, bfloat16 source dtype, 30 text layers, 16 attention heads, 8 local KV heads, 2 global KV heads, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, resident vision config, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B IT non-QAT config comparison", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of the checked text, vision, context, tied-embedding, expert, and attention geometry fields found no differences between the QAT unquantized config and the already audited non-QAT Google instruction-tuned config. The only observed difference in the fetched files is Transformers metadata." }, { "label": "Google Gemma 4 26B A4B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/tree/dfc00409adc70be497fee9c90bfe76b3ee130f2e", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-26B_q4_0-it.gguf is 14439361440 bytes, exactly matching API gguf.totalFileSize. The sidecar gemma-4-26B-it-mmproj.gguf is 1194827744 bytes and is not the selected main text artifact." }, { "label": "Google Gemma 4 26B A4B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/resolve/dfc00409adc70be497fee9c90bfe76b3ee130f2e/gemma-4-26B_q4_0-it.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 46 metadata entries and 658 tensors. The selected file is 14.439361440 GB, with tensor payloads starting at byte 15821792. Tensor spans total 14.423538808 GB across 25.233142046B logical elements: Q4_0 13.771929600 GB, Q6_K 0.605552640 GB, and F32 0.046056568 GB. Metadata/tokenizer/header/file overhead accounts for 0.015822632 GB. Non-expert and always-on tensor spans total 1.577172088 GB. Routed expert Q4_0 tensors total 12.846366720 GB across 30 layers and 128 expert indexes, or 0.100362240 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, unified full-attention K/V geometry, and separate sliding-layer K/V projections." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, comparison against the existing audited Gemma 4 26B A4B IT config/profile, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for Google's official main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--gemma-4-26b-a4b-it", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-26B-A4B-it", "title": "Google Gemma 4 26B A4B IT BF16", "summary": "Audited memory-side bounds profile for the BF16 Gemma 4 26B A4B instruction-tuned repo.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B", "relation": "finetune", "source": "Hugging Face model card base_model metadata", "config_compatible": true, "notes": "The instruction-tuned repo keeps the same text architecture as the base Gemma 4 26B A4B model." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 51.611872412, "main_resident_weight_gb": 50.46628358, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 4.79031302, "routed_expert_weight_gb": 0.35684352, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed experts", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The config records top_k_experts 8 and the public card describes 1 shared expert. The safetensors headers expose dense model.language_model.layers.*.mlp.* tensors in addition to routed model.language_model.layers.*.experts.* tensors; this profile charges those shared/always-on MLP tensors in fixed_weight_gb.", "notes": "The HF API and index metadata total parameter count is larger than the stored BF16 tensor count because the tied embedding/output projection is counted separately in metadata. The profile uses exact safetensors header bytes for resident and traffic weights. All 30 routed expert layers have uniform stacked expert tensor byte counts divisible by 128 experts." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This BF16 profile is intentionally separate from the NVIDIA NVFP4 Gemma profile used by the worked examples." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the base Google repo.", "notes": "The config dtype is bfloat16." }, "evidence": [ { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records dtype bfloat16, tie_word_embeddings true, attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, intermediate_size 2112, moe_intermediate_size 704, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B IT Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "weight_format", "total_params_b" ], "notes": "The API response at commit 20da991ab4afab98e8f910c4a2e8f4fbefc404ad records safetensors parameters BF16: 25805936206 and total: 26544131376. The direct safetensors headers have no separate lm_head.weight and sum to 25805936206 stored BF16 parameters, so exact memory uses the header/index byte count rather than the metadata total parameter count." }, { "label": "Google Gemma 4 26B A4B base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text, vision, serving, expert, and context architecture fields match the instruction-tuned repo config." }, { "label": "Google Gemma 4 26B A4B IT safetensors index and shard headers", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "tie_word_embeddings" ], "notes": "The index records total_size 51611872412 bytes across two shards. Range-read safetensors headers found 1013 BF16 tensors, no separate lm_head.weight, and stored tensors totaling 25805936206 params / 51.611872412 GB. Resident text tensors under model.language_model total 50.46628358 GB; resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic, defined as all model.language_model tensors except model.language_model.layers.*.experts.down_proj and model.language_model.layers.*.experts.gate_up_proj, totals 4.79031302 GB. Routed expert tensors total 45.67597056 GB and divide exactly into 128 uniform expert groups of 0.35684352 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the Google model card, served config, base config comparison, HF API metadata, safetensors index, and direct shard header byte grouping." }, "notes": "This profile answers direct BF16 comparisons against google/gemma-4-26B-A4B-it. Use nvidia/Gemma-4-26B-A4B-NVFP4 for the NVFP4 worked-example bounds." }, { "id": "google--gemma-4-26b-a4b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-26B-A4B", "title": "Google Gemma 4 26B A4B BF16", "summary": "Audited memory-side bounds profile for the BF16 Gemma 4 26B A4B base repo.", "model_family": "gemma4-moe", "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 51.611872412, "main_resident_weight_gb": 50.46628358, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 4.79031302, "routed_expert_weight_gb": 0.35684352, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed experts", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The config records top_k_experts 8 and the public card describes 1 shared expert. The safetensors headers expose dense model.language_model.layers.*.mlp.* tensors in addition to routed model.language_model.layers.*.experts.* tensors; this profile charges those shared/always-on MLP tensors in fixed_weight_gb.", "notes": "The HF API and index metadata total parameter count is larger than the stored BF16 tensor count because the tied embedding/output projection is counted separately in metadata. The profile uses exact safetensors header bytes for resident and traffic weights. All 30 routed expert layers have uniform stacked expert tensor byte counts divisible by 128 experts." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This BF16 profile is intentionally separate from the NVIDIA NVFP4 Gemma profile used by the worked examples." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the base Google repo.", "notes": "The config dtype is bfloat16." }, "evidence": [ { "label": "Google Gemma 4 26B A4B model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B config", "url": "https://huggingface.co/google/gemma-4-26B-A4B/raw/f1102d7de421741c6eafcda46d1806a7a65b83a3/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records dtype bfloat16, tie_word_embeddings true, attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, intermediate_size 2112, moe_intermediate_size 704, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "total_params_b" ], "notes": "The API response at commit f1102d7de421741c6eafcda46d1806a7a65b83a3 records downloads 720641, safetensors parameters BF16: 25805936206, and total: 26544131376. The direct safetensors headers have no separate lm_head.weight and sum to 25805936206 stored BF16 parameters, so exact memory uses the header/index byte count rather than the metadata total parameter count." }, { "label": "Google Gemma 4 26B A4B safetensors index and shard headers", "url": "https://huggingface.co/google/gemma-4-26B-A4B/raw/f1102d7de421741c6eafcda46d1806a7a65b83a3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "tie_word_embeddings" ], "notes": "The index records total_size 51611872412 bytes across two shards. Range-read safetensors headers found 1013 BF16 tensors, no separate lm_head.weight, and stored tensors totaling 25805936206 params / 51.611872412 GB. Resident text tensors under model.language_model total 50.46628358 GB; resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic, defined as all model.language_model tensors except model.language_model.layers.*.experts.down_proj and model.language_model.layers.*.experts.gate_up_proj, totals 4.79031302 GB. Routed expert tensors total 45.67597056 GB and divide exactly into 128 uniform expert groups of 0.35684352 GB." }, { "label": "Google Gemma 4 26B A4B IT config and shard-header comparison", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "derived_calculation", "supports": [ "architecture", "weight_adapter", "kv_adapter" ], "notes": "Manual comparison against the already audited instruction-tuned repo found no differences in relevant text, vision, serving, expert, context, tensor-name, tensor-shape, or tensor-byte fields." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from the Google model card, served config, HF API metadata, safetensors index, direct shard header byte grouping, and comparison against the already audited instruction-tuned sibling." }, "notes": "This profile answers direct BF16 comparisons against google/gemma-4-26B-A4B. Use nvidia/Gemma-4-26B-A4B-NVFP4 for the NVFP4 worked-example bounds." }, { "id": "google--gemma-4-31b-it-assistant", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-4-31B-it-assistant", "title": "Google Gemma 4 31B IT Assistant Drafter BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 Gemma 4 31B IT MTP assistant drafter repo.", "model_family": "gemma4-assistant-speculative-drafter", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "derived_package", "source": "Model card speculative-decoding instructions and generation_config assistant metadata", "config_compatible": false, "notes": "The model card identifies this repo as the Multi-Token Prediction drafter for google/gemma-4-31B-it. It is loaded as assistant_model alongside the target 31B model in Transformers speculative decoding, not as a replacement target model." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-assistant-drafter", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.469518596, "swept_params_b": 0.469518596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.939037192, "swept_weight_gb": 0.939037192, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "Exact BF16 tensor bytes are recorded, but Bounds Engine v1 does not use them for production throughput because this repo is a speculative drafter paired with a target model.", "auxiliary_scope": "No separate auxiliary resident tensors were identified in the single safetensors file; the repo stores model.embed_tokens, model.layers, model.norm, pre_projection, and post_projection tensors.", "notes": "The single safetensors file stores 48 BF16 tensors totaling 0.939037192 GB: model.embed_tokens 0.536870912 GB, model.layers 0.369134088 GB, model.norm 0.000002048 GB, pre_projection.weight 0.022020096 GB, and post_projection.weight 0.011010048 GB." }, "kv_adapter": { "kind": "unknown", "reason": "Gemma4AssistantForCausalLM is an MTP speculative drafter. Production throughput depends on the paired target model, draft length, acceptance rate, verification schedule, and extra target-side validation work. Bounds Engine v1 only models single-model ordinary autoregressive text-token decode.", "notes": "The config records four assistant text layers with three sliding-attention layers and one full-attention layer, but using those fields as ordinary target-model decode would be misleading because generation_config marks is_assistant true and num_assistant_tokens 6." }, "notes": "This profile intentionally fails closed even though exact weights and assistant-layer geometry are accessible, because comparison tok/s should not treat this drafter as a standalone target model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-gemma4-mtp-assistant-drafter", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo is used as assistant_model in speculative decoding with google/gemma-4-31B-it. A production profile needs a paired target+assistant speculative-decoding adapter, not a standalone ordinary text-decode adapter." }, "evidence": [ { "label": "Google Gemma 4 31B IT Assistant API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-assistant", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 34ef9f029d1c52bccac2def222523af32f3ccd0f, the API reports an Apache-2.0 any-to-any repo with transformers, safetensors, gemma4_assistant, text-generation, endpoints_compatible, region:us, 806036 downloads, and BF16 safetensors metadata for 469518596 parameters." }, { "label": "Google Gemma 4 31B IT Assistant config", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/raw/34ef9f029d1c52bccac2def222523af32f3ccd0f/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "assistant_geometry", "unsupported_reason" ], "notes": "The config records Gemma4AssistantForCausalLM, model_type gemma4_assistant, BF16 dtype, backbone_hidden_size 5376, 2048 centroids, centroid_intermediate_top_k 32, tie_word_embeddings true, and a four-layer text_config with three sliding-attention layers, one full-attention layer, hidden_size 1024, intermediate_size 8192, 32 attention heads, 16 local KV heads, 4 global KV heads, attention_k_eq_v true, 1024-token sliding window, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT Assistant generation config", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/raw/34ef9f029d1c52bccac2def222523af32f3ccd0f/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config explicitly records is_assistant true, num_assistant_tokens 6, and a constant assistant-token schedule." }, { "label": "Google Gemma 4 31B IT Assistant model card", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm", "base_model_proof" ], "notes": "The model card states that this checkpoint contains Multi-Token Prediction drafters for Gemma 4 models. Its example loads google/gemma-4-31B-it as TARGET_MODEL_ID and this repo as ASSISTANT_MODEL_ID, then passes assistant_model=assistant_model to target_model.generate." }, { "label": "Google Gemma 4 31B IT Assistant safetensors header", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/resolve/34ef9f029d1c52bccac2def222523af32f3ccd0f/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "resident_weight_gb", "weight_format" ], "notes": "A range-read of the single safetensors header found a 5360-byte header and 48 BF16 tensors totaling 469518596 parameters / 0.939037192 GB. Tensor groups are model.embed_tokens 0.536870912 GB, model.layers 0.369134088 GB, model.norm 0.000002048 GB, pre_projection.weight 0.022020096 GB, and post_projection.weight 0.011010048 GB. The linked file size is 939042560 bytes, so non-tensor safetensors overhead is 5368 bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from live HF API metadata, model card, served config, generation config, and direct safetensors header byte grouping. Marked unsupported because Bounds Engine v1 lacks a paired target+assistant speculative decoding adapter." }, "unsupported_reason": "Gemma 4 assistant checkpoints are speculative drafters used with a separate target model. Bounds Engine v1 does not model paired target+assistant speculative decoding, acceptance rates, or target verification work, so standalone tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports paired speculative-decoding profiles." }, { "id": "google--gemma-4-31b-it-qat-q4-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-31B-it-qat-q4_0-gguf", "title": "Google Gemma 4 31B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for Google's official Q4_0 GGUF artifact of Gemma 4 31B IT QAT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, model card architecture table, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-31B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 31B dense text geometry as the Google QAT unquantized config. The Google official GGUF repo ships only GGUF files, so the immutable Google QAT unquantized config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 17.650999456, "swept_weight_gb": 17.635168128, "auxiliary_resident_weight_gb": 0.015831328, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31B_q4_0-it.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as model tensors; gemma-4-31B-it-mmproj.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_0 GGUF file selected by HF API gguf.totalFileSize. Header tensor payloads total 17.635166448 GB, with 0.000001680 GB of tensor-alignment padding, while the linked file size is 17.650999456 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config and GGUF metadata record four global KV heads and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5750008384536024, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is Google's official Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-31B_q4_0-it.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Google Gemma 4 31B IT QAT Q4_0 GGUF HF API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-qat-q4_0-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 4a311c5261daa0702f80836f8866114943651ab0 records base_model google/gemma-4-31B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, image-text-to-text pipeline, region:us, 284997 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, and gguf.totalFileSize 17650999456." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-gguf/raw/4a311c5261daa0702f80836f8866114943651ab0/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "attention_pattern", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-31B-it-qat-q4_0-unquantized, QAT release packaging, Q4_0 GGUF ready-to-deploy artifacts, Gemma 4 31B dense architecture, 30.7B parameters, 60 layers, 1024-token sliding window, 256K context, 262K vocabulary, text/image modalities, and about 550M vision encoder parameters. It also describes hybrid local/global attention with unified K/V global layers." }, { "label": "Google Gemma 4 31B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, four global KV heads, 512 global head dimension, 16 local KV heads, 256 local head dimension, dense MLPs, resident Gemma 4 vision config, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-gguf/tree/4a311c5261daa0702f80836f8866114943651ab0", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-31B_q4_0-it.gguf is 17650999456 bytes, exactly matching API gguf.totalFileSize. The sidecar gemma-4-31B-it-mmproj.gguf is 1200726016 bytes and is not the selected main text artifact." }, { "label": "Google Gemma 4 31B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-gguf/resolve/4a311c5261daa0702f80836f8866114943651ab0/gemma-4-31B_q4_0-it.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 43 metadata entries and 833 tensors. The selected file is 17.650999456 GB, with tensor payloads starting at byte 15831328. Tensor spans total 17.635168128 GB across 30.697345596B logical elements: token_embd.weight 1.156055040 GB, blk.* tensors 16.473784320 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Tensor payloads split into Q4_0 16.473784320 GB, Q6_K 1.156055040 GB, and F32 0.005327088 GB. Metadata/tokenizer/header/file overhead accounts for 0.015831328 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, a layer KV head array with ten global layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for Google's official main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--gemma-4-31b-it-qat-q4-0-unquantized-assistant", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-4-31B-it-qat-q4_0-unquantized-assistant", "title": "Google Gemma 4 31B IT QAT Q4_0 Unquantized Assistant Drafter BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 Gemma 4 31B IT QAT unquantized MTP assistant drafter repo.", "model_family": "gemma4-assistant-speculative-drafter", "base_model_proof": { "base_model": "google/gemma-4-31B-it-assistant", "relation": "derived_package", "source": "Hugging Face model card/API metadata, served config, Transformers Gemma4Assistant implementation, and direct safetensors header evidence", "config_compatible": false, "notes": "The repo card identifies this package as derived from google/gemma-4-31B-it-assistant. Its served config uses Gemma4AssistantForCausalLM with the same 31B assistant geometry as the base assistant repo. Transformers documents and implements Gemma4Assistant as a Multi-Token Prediction speculative drafter that is loaded as assistant_model with a separate target Gemma 4 model, not as a replacement target model." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-qat-q4-0-unquantized-assistant-drafter", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.469518596, "swept_params_b": 0.469518596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.939037192, "swept_weight_gb": 0.939037192, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "Exact BF16 tensor bytes are recorded, but Bounds Engine v1 does not use them for production throughput because this repo is a speculative drafter paired with a target model.", "auxiliary_scope": "No separate auxiliary resident tensors were identified in the single safetensors file; the repo stores model.embed_tokens, model.layers, model.norm, pre_projection, and post_projection tensors.", "notes": "The single safetensors file stores 48 BF16 tensors totaling 0.939037192 GB: model.embed_tokens 0.536870912 GB, model.layers 0.369134088 GB, model.norm 0.000002048 GB, pre_projection.weight 0.022020096 GB, and post_projection.weight 0.011010048 GB. The linked file size is 0.939042560 GB, so non-tensor safetensors overhead is 0.000005368 GB." }, "kv_adapter": { "kind": "unknown", "reason": "Gemma4AssistantForCausalLM is an MTP speculative drafter. Production throughput depends on the paired target model, draft length, acceptance rate, verification schedule, shared target KV state, and extra target-side validation work. Bounds Engine v1 only models single-model ordinary autoregressive text-token decode.", "notes": "The config records four assistant text layers with three sliding-attention layers and one full-attention layer, but using those fields as ordinary target-model decode would be misleading because the implementation requires target-model inputs_embeds and shared_kv_states." }, "notes": "This profile intentionally fails closed even though exact weights and assistant-layer geometry are accessible, because comparison tok/s should not treat this drafter as a standalone target model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-gemma4-mtp-assistant-drafter", "dequantization_notes": "No quantized weight representation is assumed for this BF16 unquantized-QAT assistant repo.", "notes": "The repo is used as an assistant_model derived from google/gemma-4-31B-it-assistant. A production profile needs a paired target+assistant speculative-decoding adapter, not a standalone ordinary text-decode adapter." }, "evidence": [ { "label": "Google Gemma 4 31B QAT Q4_0 Unquantized Assistant API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "base_model_proof" ], "notes": "At commit 2194ebe30f476ea96d5e65503695dea6d78791bf, the API reports an Apache-2.0 public non-gated image-text-to-text repo with transformers, safetensors, gemma4_assistant, base_model google/gemma-4-31B-it-assistant, endpoints_compatible, region:us, 143335 downloads, and BF16 safetensors metadata for 469518596 parameters." }, { "label": "Google Gemma 4 31B QAT Q4_0 Unquantized Assistant config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/raw/2194ebe30f476ea96d5e65503695dea6d78791bf/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "assistant_geometry", "unsupported_reason" ], "notes": "The config records Gemma4AssistantForCausalLM, model_type gemma4_assistant, BF16 dtype, backbone_hidden_size 5376, 2048 centroids, centroid_intermediate_top_k 32, tie_word_embeddings true, and a four-layer text_config with three sliding-attention layers, one full-attention layer, hidden_size 1024, intermediate_size 8192, 32 attention heads, 16 local KV heads, 4 global KV heads, attention_k_eq_v true, 1024-token sliding window, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B QAT Q4_0 Unquantized Assistant generation config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/raw/2194ebe30f476ea96d5e65503695dea6d78791bf/generation_config.json", "source_type": "config", "supports": [ "generation_settings" ], "notes": "The generation config records normal sampling defaults but does not by itself mark the repo as an assistant. Assistant status is established by the Gemma4AssistantForCausalLM class, model card lineage, and Transformers assistant documentation." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant", "source_type": "model_card", "supports": [ "unsupported_reason", "base_model_proof", "qat_unquantized_scope" ], "notes": "The card describes unquantized QAT checkpoints as half-precision weights extracted from the QAT pipeline, including drafter models, and identifies this package's base model as google/gemma-4-31B-it-assistant." }, { "label": "Transformers Gemma4Assistant implementation", "url": "https://github.com/huggingface/transformers/blob/08a7ef05bcf9723cb2e58855afb8dc2c799323ff/src/transformers/models/gemma4_assistant/modeling_gemma4_assistant.py", "source_type": "manual_review", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "Local review of the Hugging Face Transformers checkout at 08a7ef05bcf9723cb2e58855afb8dc2c799323ff found Gemma4AssistantForCausalLM documented as a model for multi-token prediction assisted decoding. Its forward method raises unless inputs_embeds and shared_kv_states are provided, projects target-model embedding/hidden-state inputs into assistant space, and cross-attends to shared target-model KV states." }, { "label": "Transformers Gemma4Assistant documentation", "url": "https://github.com/huggingface/transformers/blob/08a7ef05bcf9723cb2e58855afb8dc2c799323ff/docs/source/en/model_doc/gemma4_assistant.md", "source_type": "manual_review", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The Transformers documentation describes Gemma 4 Assistant as a small text-only model for speculative decoding with Gemma 4 using Multi-Token Prediction. Its examples load a target Gemma 4 model and a separate assistant_model, then call target_model.generate(..., assistant_model=assistant_model)." }, { "label": "Google Gemma 4 31B QAT Q4_0 Unquantized Assistant safetensors header", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/resolve/2194ebe30f476ea96d5e65503695dea6d78791bf/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "resident_weight_gb", "weight_format" ], "notes": "A range-read of the single safetensors header found a 5360-byte header and 48 BF16 tensors totaling 469518596 parameters / 0.939037192 GB. Tensor groups are model.embed_tokens 0.536870912 GB, model.layers 0.369134088 GB, model.norm 0.000002048 GB, pre_projection.weight 0.022020096 GB, and post_projection.weight 0.011010048 GB. The linked file size is 939042560 bytes, so non-tensor safetensors overhead is 5368 bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from current HF API metadata, model card, served config, generation config, direct safetensors header byte grouping, and local review of the Hugging Face Transformers Gemma4Assistant implementation/documentation. Marked unsupported because Bounds Engine v1 lacks a paired target+assistant speculative decoding adapter." }, "unsupported_reason": "Gemma 4 assistant checkpoints are speculative drafters used with a separate target model. Bounds Engine v1 does not model paired target+assistant speculative decoding, shared target KV state, acceptance rates, or target verification work, so standalone tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports paired speculative-decoding profiles." }, { "id": "google--gemma-4-31b-it-qat-w4a16-ct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-31B-it-qat-w4a16-ct", "title": "Google Gemma 4 31B IT QAT W4A16 CT", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors W4A16 QAT Gemma 4 31B instruction-tuned repo.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The compressed-tensors repo records google/gemma-4-31B-it-qat-q4_0-unquantized as its base model, and that unquantized QAT repo records google/gemma-4-31B-it as its base. Manual comparison found matching text and vision decode geometry between the compressed and unquantized QAT configs. The compressed repo adds quantization_config plus explicit top-level hidden_size and intermediate_size defaults that do not change the text-decode fields used by this profile." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.59758608, "swept_params_b": 31.6125564, "auxiliary_resident_params_b": 1.98502968, "resident_weight_gb": 23.26508556, "swept_weight_gb": 19.2950262, "auxiliary_resident_weight_gb": 3.97005936, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens input table, model.vision_tower, and model.embed_vision tensors are resident for the package but not full-matrix swept for each generated text token", "notes": "Compressed-tensors stores packed I32 int4 weights, BF16 scale and ignored-module tensors, and tiny I64 weight_shape side tensors. The config records tie_word_embeddings true, but this compressed artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection and treats model.language_model.embed_tokens.weight as a resident input table for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config records attention_k_eq_v true, num_global_key_value_heads 4, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Quantizing weights does not change the BF16 KV cache assumption because the config has kv_cache_scheme null." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6924630092353349, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scale tensors, ignored BF16 modules, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, bfloat16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records the exact resident stored-byte average across logical safetensors parameters; resident and swept traffic use the exact header byte fields above." }, "evidence": [ { "label": "Google Gemma 4 31B IT QAT W4A16 CT model card and API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-qat-w4a16-ct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The live HF API response records repo SHA ca9d5250d08695be35f2d632b434938b6d4aecfe, Apache-2.0 licensing, image-text-to-text pipeline, region:us tag, base_model google/gemma-4-31B-it-qat-q4_0-unquantized, compressed-tensors tags, 987445 downloads, and safetensors logical parameters I64: 820, I32: 29286727680, BF16: 4310857580, total: 33597586080. The model card describes the W4A16 compressed-tensors artifacts as QAT checkpoints serialized for native optimized inference with vLLM." }, { "label": "Google Gemma 4 31B IT QAT W4A16 CT config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct/raw/ca9d5250d08695be35f2d632b434938b6d4aecfe/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, compressed-tensors pack-quantized int4 weights, group_size 32, symmetric group strategy, quantization_status compressed, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, num_global_key_value_heads 4, global_head_dim 512, 16 sliding KV heads, 256 sliding head dimension, 262144 max position embeddings, and Gemma 4 vision config. The quantization ignore list has 192 entries covering vision modules, model.embed_vision.embedding_projection, and lm_head." }, { "label": "Google Gemma 4 31B IT QAT unquantized base config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching audited text_config and vision_config geometry fields between the compressed-tensors repo and the unquantized QAT base. The compressed repo adds quantization_config and explicit top-level hidden_size/intermediate_size defaults; these do not change ordinary text-decode architecture fields." }, { "label": "Google Gemma 4 31B IT QAT W4A16 CT safetensors header", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct/resolve/ca9d5250d08695be35f2d632b434938b6d4aecfe/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The linked object size is 23265352448 bytes, with a 266880-byte header and 23265085560 tensor bytes across 2009 tensors. Stored tensors sum to 33.59758608B logical parameters / 23.26508556 GB: 410 I32 packed tensors carrying 29.28672768B logical int4 weights / 14.64336384 GB, 1189 BF16 tensors carrying 4.31085758B parameters / 8.62171516 GB, and 410 I64 shape tensors carrying 820 values / 0.00000656 GB. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens.weight plus lm_head.weight, sum to 31.6125564B logical parameters / 19.2950262 GB. Resident-only tensors, defined as model.language_model.embed_tokens.weight plus model.vision_tower and model.embed_vision tensors, sum to 1.98502968B parameters / 3.97005936 GB. Full-attention layer headers have no v_proj tensors; sliding-attention layer headers have separate v_proj tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, the served config, unquantized QAT base config comparison, model card, linked object metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident input embedding and multimodal tensors from per-token swept language/logit weights." }, { "id": "google--gemma-4-31b-it", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-31B-it", "title": "Google Gemma 4 31B IT BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 31B instruction-tuned repo.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B", "relation": "finetune", "source": "Hugging Face model card base_model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text tensor geometry, context fields, attention pattern, dtype, and dense/MoE setting between the instruction-tuned and base configs." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 31.273088876, "swept_params_b": 30.69734534, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 62.546177752, "swept_weight_gb": 61.39469068, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for each generated text token", "notes": "The HF API total_parameters value is larger than stored BF16 parameters. This profile uses safetensors header stored BF16 counts for memory footprint and language-model tensor headers for per-token decode traffic. The config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text and vision configs record bfloat16, and safetensors headers record BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 31B IT model card", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model metadata identifies google/gemma-4-31B as the base model, Apache-2.0 licensing, and image-text-to-text packaging." }, { "label": "Google Gemma 4 31B IT config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, and a resident vision config." }, { "label": "Google Gemma 4 31B IT Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 3548789868c5356dbf307c98e6f609007b82b3eb records safetensors parameters BF16: 31273088876 and total: 32682372656." }, { "label": "Google Gemma 4 31B base config", "url": "https://huggingface.co/google/gemma-4-31B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text architecture and context fields match the instruction-tuned repo config." }, { "label": "Google Gemma 4 31B IT safetensors index", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across both shards. Stored tensors sum to 31273088876 BF16 params and 62.546177752 GB. model.language_model.embed_tokens.weight is 1409286144 BF16 params and 2.818572288 GB. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 30697345340 BF16 params and 61.39469068 GB. Auxiliary resident tensors, defined as model.vision_tower plus model.embed_vision tensors, sum to 575743536 BF16 params and 1.151487072 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API metadata, and safetensors header parameter split." }, "notes": "This profile is for text decode bounds. It deliberately separates resident multimodal weights from per-token swept language weights." }, { "id": "google--gemma-4-31b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-31B", "title": "Google Gemma 4 31B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 31B base repo.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B", "relation": "base", "source": "Hugging Face model metadata, served config, direct safetensors header grouping, and instruction-tuned sibling comparison", "config_compatible": true, "notes": "This profile targets the base BF16 repo directly. Manual comparison with the already audited google/gemma-4-31B-it sibling found matching text tensor geometry, context fields, attention pattern, dtype, dense/MoE setting, and relevant vision geometry; the base repo config is used as authoritative here." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 31.273088876, "swept_params_b": 30.69734534, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 62.546177752, "swept_weight_gb": 61.39469068, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "all model.language_model safetensors headers, including tied model.language_model.embed_tokens.weight as output projection traffic", "auxiliary_scope": "model.vision_tower and model.embed_vision tensors are resident for the multimodal package but not swept for each generated text token", "notes": "The HF API total parameter value is larger than stored BF16 parameters because metadata counts the tied output projection separately. This profile uses safetensors header stored BF16 counts for memory footprint and language-model tensor headers for per-token decode traffic. The config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text and vision configs record bfloat16, and safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 31B model card and API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b", "max_context_tokens" ], "notes": "At commit 02e15e4990e8c452f8543fb26beff15b1daf8f3d, the API records a public Apache-2.0 image-text-to-text BF16 repo with current downloads 656084, region:us, and safetensors parameters BF16 31273088876 / total 32682372656. The model card describes Gemma 4 31B as a dense text/image model with 256K context, 1024-token sliding window, hybrid local/global attention, and about 550M vision encoder parameters." }, { "label": "Google Gemma 4 31B config", "url": "https://huggingface.co/google/gemma-4-31B/raw/02e15e4990e8c452f8543fb26beff15b1daf8f3d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 32 attention heads, 16 sliding KV heads, 4 global KV heads, 256 local head dimension, 512 global key/value length, 262144 max position embeddings, and a resident Gemma 4 vision config." }, { "label": "Google Gemma 4 31B IT config comparison", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison against the already audited instruction-tuned sibling found no differences in relevant text_config fields. Top-level differences are limited to generation-token metadata, and non-geometry vision config defaults do not change the text-decode profile." }, { "label": "Google Gemma 4 31B safetensors index and shard headers", "url": "https://huggingface.co/google/gemma-4-31B/resolve/02e15e4990e8c452f8543fb26beff15b1daf8f3d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "tie_word_embeddings" ], "notes": "Safetensors headers were range-read across both indexed shards. Stored tensors sum exactly to the index total_size, 62.546177752 GB, across 1188 BF16 tensors. model.language_model tensors sum to 61.39469068 GB and are swept for ordinary text decode, including the tied model.language_model.embed_tokens.weight tensor of 2.818572288 GB. Auxiliary resident tensors, defined as model.vision_tower plus model.embed_vision tensors, sum to 1.151487072 GB. The index has no separate lm_head.weight. All index weight_map entries were found in the shard headers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, served config, instruction-tuned sibling config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile is for text decode bounds. It deliberately separates resident multimodal weights from per-token swept language weights." }, { "id": "google--gemma-4-e2b-it-qat-q4-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E2B-it-qat-q4_0-gguf", "title": "Google Gemma 4 E2B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for Google's official Q4_0 GGUF artifact of Gemma 4 E2B IT QAT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, Google Gemma 4 E2B IT config comparison, Transformers PLE documentation, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E2B-it-qat-q4_0-unquantized. Manual comparison found the QAT unquantized config and the audited Gemma 4 E2B IT config have matching text, multimodal, context, attention, PLE, and serving-relevant fields; the only checked text_config difference is a null field renamed from expert_intermediate_size to moe_intermediate_size." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.628569635, "swept_params_b": 2.279759395, "auxiliary_resident_params_b": 2.34881024, "resident_weight_gb": 3.349514112, "swept_weight_gb": 1.406942304, "auxiliary_resident_weight_gb": 1.942571808, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-E2B_q4_0-it.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as full matrices per generated token; gemma-4-E2B-it-mmproj.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "Gemma 4 E2B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per_layer_proj_norm.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. The Google config records num_kv_shared_layers 20, so only layers 4, 9, and 14 allocate full-context K/V cache, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records shared_kv_layers 20. Only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7236607367148318, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-ple-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is Google's official Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-E2B_q4_0-it.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Google Gemma 4 E2B IT QAT Q4_0 GGUF HF API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B-it-qat-q4_0-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 1894d1fc0a19d86697abd40483f5983c867df03f records base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, any-to-any pipeline, region:us, 346498 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 4628569635, and gguf.totalFileSize 3349514112." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "model_family", "per_layer_embeddings", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, QAT release packaging, Q4_0 GGUF ready-to-deploy artifacts, any-to-any multimodal support, 128K context for E2B/E4B, 35 E2B layers, 512-token sliding window, 262K vocabulary, text/image/audio modalities, and Per-Layer Embeddings for E2B/E4B." }, { "label": "Google Gemma 4 E2B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, one KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E2B IT config comparison", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter", "per_layer_embeddings" ], "notes": "Manual comparison against the already audited Gemma 4 E2B IT config found matching checked top-level text, vision, audio, dtype, context, attention, PLE, and serving-relevant fields. The only checked text_config difference is a null field name: expert_intermediate_size in the base config versus moe_intermediate_size in the QAT unquantized config." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "Google Gemma 4 E2B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf/tree/1894d1fc0a19d86697abd40483f5983c867df03f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E2B_q4_0-it.gguf is 3349514112 bytes, exactly matching API gguf.totalFileSize. The sidecar gemma-4-E2B-it-mmproj.gguf is 986833312 bytes and is not the selected main text artifact." }, { "label": "Google Gemma 4 E2B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf/resolve/1894d1fc0a19d86697abd40483f5983c867df03f/gemma-4-E2B_q4_0-it.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 43 metadata entries and 541 tensors. The selected file is 3.349514112 GB, with tensor payloads starting at byte 15813408. Tensor spans total 3.333700704 GB across 4628569635 logical elements: per_layer_token_embd.weight 1.926758400 GB, token_embd.weight 0.330301440 GB, blk.* tensors 1.049107552 GB, per_layer_model_proj/proj_norm tensors 0.027526144 GB, output_norm.weight 0.000006144 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/file overhead accounts for 0.015813408 GB. Swept tensor spans excluding the per-layer token lookup table total 1.406942304 GB across 2279759395 logical elements. Tensor payloads split into Q4_0 1.047969792 GB, Q6_K 2.257059840 GB, F16 0.027525120 GB, and F32 0.001144972 GB, with 980 bytes of tensor-alignment padding. The header records gemma4.block_count 35, context_length 131072, attention.head_count 8, one KV head, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, shared_kv_layers 20, embedding_length_per_layer_input 256, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, comparison against the existing audited Gemma 4 E2B IT config/profile, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for Google's official main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--gemma-4-e2b-it-qat-w4a16-ct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E2B-it-qat-w4a16-ct", "title": "Google Gemma 4 E2B IT QAT W4A16 CT", "summary": "Audited memory-side text-decode bounds profile for Google's compressed-tensors W4A16 QAT Gemma 4 E2B IT repo.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API metadata, served compressed-tensors config, Google QAT unquantized config, existing Gemma 4 E2B profile comparison, Transformers Gemma 4 implementation, and direct safetensors header metadata", "config_compatible": true, "notes": "The compressed-tensors repo records google/gemma-4-E2B-it-qat-q4_0-unquantized as its base model. Manual comparison found the compressed, QAT unquantized, and already audited Gemma 4 E2B IT configs match on the checked text, multimodal, context, attention, PLE, and serving-relevant fields; the compressed repo adds quantization_config and a null text_config field renamed from expert_intermediate_size to moe_intermediate_size." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.565602827, "swept_params_b": 2.338410315, "auxiliary_resident_params_b": 3.227192512, "resident_weight_gb": 8.315979014, "swept_weight_gb": 1.86159399, "auxiliary_resident_weight_gb": 6.454385024, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "model.language_model excluding embed_tokens and embed_tokens_per_layer plus separate lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens input table, model.language_model.embed_tokens_per_layer PLE lookup table, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision tensors are resident for the multimodal PLE package but not full-matrix swept for each generated text token", "notes": "Compressed-tensors stores packed I32 int4 weights, BF16 scale and ignored-module tensors, and tiny I64 weight_shape side tensors. The config records tie_word_embeddings true, but this compressed artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection while treating model.language_model.embed_tokens.weight and model.language_model.embed_tokens_per_layer.weight as resident lookup/input tables for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. Quantizing weights does not change the BF16 KV cache assumption because the compressed-tensors config has kv_cache_scheme null." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary cached text decode after any multimodal prefill, not encoder or input-projection throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.4941739956105566, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-ple-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scale tensors, ignored BF16 modules, BF16 embeddings/towers/lm_head, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, bfloat16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records the exact resident stored-byte average across header-logical safetensors parameters; resident and swept traffic use the exact header byte fields above." }, "evidence": [ { "label": "Google Gemma 4 E2B IT QAT W4A16 CT API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B-it-qat-w4a16-ct", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The live HF API response at commit a50d96741aa53e1e0f1e8b6ea73230dce02b3341 records a public non-gated Apache-2.0 any-to-any repo with transformers, safetensors, compressed-tensors, endpoints_compatible, region:us, base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, 381034 downloads, and API safetensors logical parameters I64: 552, I32: 1876819968, BF16: 3688781379, total: 5565601899. The direct header has 928 additional BF16 elements, so this profile uses direct header counts for exact memory bytes." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family", "per_layer_embeddings", "max_context_tokens", "weight_format" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, QAT release packaging, compressed-tensors W4A16 artifacts for native optimized inference with vLLM, any-to-any multimodal support, 128K context for E2B, 35 E2B layers, 512-token sliding window, 262K vocabulary, text/image/audio modalities, about 150M vision parameters, about 300M audio parameters, and Per-Layer Embeddings for E2B/E4B." }, { "label": "Google Gemma 4 E2B IT QAT W4A16 CT config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct/raw/a50d96741aa53e1e0f1e8b6ea73230dce02b3341/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings", "kv_sharing" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, compressed-tensors pack-quantized int4 weights, group_size 32, symmetric group strategy, quantization_status compressed, kv_cache_scheme null, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, one KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config. The explicit quantization ignore list includes lm_head, audio/embed/vision projection tensors, and vision/audio tower modules." }, { "label": "Google Gemma 4 E2B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter", "per_layer_embeddings" ], "notes": "Manual comparison found the QAT unquantized config and this compressed-tensors config match the checked top-level text, vision, audio, dtype, context, attention, PLE, and serving-relevant fields. The only checked text_config difference against the already audited Gemma 4 E2B IT config is a null field name: expert_intermediate_size in the BF16 profile versus moe_intermediate_size here." }, { "label": "Transformers Gemma 4 implementation and PLE documentation", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py", "source_type": "manual_review", "supports": [ "kv_sharing", "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The upstream Gemma 4 implementation computes first_kv_shared_layer_idx = num_hidden_layers - num_kv_shared_layers, omits k_proj/v_proj/k_norm/v_norm for shared layers, updates past_key_values only for non-shared layers, and still passes shared key/value states to every shared attention layer. This supports charging fewer allocation layers than read layers. The Gemma 4 PLE path uses token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection, supporting resident-only treatment of the large per-layer embedding table while charging projection and block tensors as swept traffic." }, { "label": "Google Gemma 4 E2B IT QAT W4A16 CT safetensors header", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct/resolve/a50d96741aa53e1e0f1e8b6ea73230dce02b3341/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings", "kv_sharing" ], "notes": "The single safetensors file was range-read directly. The linked Content-Length is 8316306646 bytes, with a 327624-byte header and 8315979014 tensor bytes across 2504 tensors. Stored tensors sum to 5.565602827B header-logical parameters / 8.315979014 GB: 276 I32 packed tensors carrying 1.876819968B logical int4 weights / 0.938409984 GB, 1952 BF16 tensors carrying 3.688782307B parameters / 7.377564614 GB, and 276 I64 shape tensors carrying 552 values / 0.000004416 GB. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens and embed_tokens_per_layer plus lm_head, sum to 2.338410315B logical parameters / 1.861593990 GB. Resident-only tensors, defined as model.language_model.embed_tokens, model.language_model.embed_tokens_per_layer, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision tensors, sum to 3.227192512B parameters / 6.454385024 GB. Header grouping also confirms k_proj/v_proj tensors are absent from the shared KV layers, matching num_kv_shared_layers 20." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable served config, QAT unquantized config comparison, existing Gemma 4 E2B profile comparison, upstream Transformers Gemma 4 implementation review, and direct safetensors header byte grouping." }, "notes": "Use this profile for Google's official W4A16 compressed-tensors artifact. It models ordinary cached text decode after any multimodal prefill and deliberately separates resident input/PLE/multimodal tensors from per-token swept language/logit weights." }, { "id": "google--gemma-4-e2b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E2B-it", "title": "Google Gemma 4 E2B IT BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 E2B instruction-tuned repo.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B", "relation": "finetune", "source": "Hugging Face model card base_model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text, vision, audio, dtype, context, and attention fields between the instruction-tuned and base configs. The instruction-tuned config adds top-level chat/token IDs, which do not change the memory-side architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 10.246357958, "swept_weight_gb": 4.597279302, "auxiliary_resident_weight_gb": 5.649078656, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers_excluding_per_layer_embedding_table", "auxiliary_scope": "model.audio_tower, model.embed_audio, model.vision_tower, model.embed_vision, and model.language_model.embed_tokens_per_layer.weight are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "Gemma 4 E2B uses per-layer embeddings. The large model.language_model.embed_tokens_per_layer.weight table is resident and looked up during text decode, not read as a full matrix per token. The standard model.language_model.embed_tokens.weight table remains in swept traffic because it also serves the output projection in this tied-embedding repo." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level, text, vision, and audio configs record bfloat16 dtype, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 E2B IT model card", "url": "https://huggingface.co/google/gemma-4-E2B-it", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "ple_resident_split", "layers", "max_context_tokens" ], "notes": "The model metadata identifies google/gemma-4-E2B as the base model, Apache-2.0 licensing, and any-to-any multimodal packaging. The card states E2B has 2.3B effective parameters, 5.1B with embeddings, 35 layers, 512-token sliding window, 128K context, 262K vocabulary, image/audio support, about 150M vision encoder parameters, about 300M audio encoder parameters, and per-layer embeddings." }, { "label": "Google Gemma 4 E2B IT config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 text/vision/audio configs, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E2B base config", "url": "https://huggingface.co/google/gemma-4-E2B/raw/63db66a33dc06d58c02b1e887446e103c202602c/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text, multimodal, serving, and context architecture fields match the instruction-tuned repo config." }, { "label": "Google Gemma 4 E2B IT Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B-it", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "base_model_proof", "total_params_b" ], "notes": "The API response at commit 70af34e20bd4b7a91f0de6b22675850c43922a03 records safetensors parameters BF16: 5123178051 and total: 5123178051, any-to-any pipeline, Apache-2.0 licensing, and base_model google/gemma-4-E2B. The direct safetensors header sum is 928 parameters higher, so this profile uses the header count for exact memory bytes." }, { "label": "Google Gemma 4 E2B IT safetensors header", "url": "https://huggingface.co/google/gemma-4-E2B-it/resolve/70af34e20bd4b7a91f0de6b22675850c43922a03/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. The file Content-Length is 10246621918 bytes, with a 263952-byte header and 10246357958 tensor bytes. Stored tensors sum to 5123178979 BF16 parameters / 10.246357958 GB. model.language_model tensors sum to 4647449891 BF16 parameters / 9.294899782 GB. model.language_model.embed_tokens.weight has shape [262144, 1536] and contributes 0.805306368 GB. model.language_model.embed_tokens_per_layer.weight has shape [262144, 8960] and contributes 4.69762048 GB. The header has no separate lm_head.weight, so the standard embedding remains swept as the tied output projection. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens_per_layer, sum to 2298639651 BF16 parameters / 4.597279302 GB. Resident-only tensors, defined as audio tower/projection plus vision tower/projection plus the per-layer embedding table, sum to 2824539328 BF16 parameters / 5.649078656 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from served config, base config comparison, model card/API metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal and PLE weights from per-token swept language weights." }, { "id": "google--gemma-4-e2b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E2B", "title": "Google Gemma 4 E2B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 E2B base repo.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B", "relation": "base", "source": "Hugging Face model card/API metadata, base config, direct safetensors header, and instruction-tuned sibling config comparison", "config_compatible": true, "notes": "This is the base Gemma 4 E2B repo. Manual comparison against google/gemma-4-E2B-it found matching text, vision, audio, dtype, context, attention, and PLE fields; the instruction-tuned config only adds a top-level eos_token_id." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 10.246357958, "swept_weight_gb": 4.597279302, "auxiliary_resident_weight_gb": 5.649078656, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers_excluding_per_layer_embedding_table", "auxiliary_scope": "model.audio_tower, model.embed_audio, model.vision_tower, model.embed_vision, and model.language_model.embed_tokens_per_layer.weight are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "Gemma 4 E2B uses per-layer embeddings. The large model.language_model.embed_tokens_per_layer.weight table is resident and looked up during text decode, not read as a full matrix per token. The standard model.language_model.embed_tokens.weight table remains in swept traffic because it also serves the output projection in this tied-embedding repo." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level, text, vision, and audio configs record bfloat16 dtype, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 E2B model card", "url": "https://huggingface.co/google/gemma-4-E2B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "ple_resident_split", "layers", "max_context_tokens", "modalities" ], "notes": "The model metadata records Apache-2.0 licensing and any-to-any multimodal packaging. The card states E2B has 2.3B effective parameters, 5.1B with embeddings, 35 layers, 512-token sliding window, 128K context, 262K vocabulary, image/audio support, about 150M vision encoder parameters, about 300M audio encoder parameters, and per-layer embeddings." }, { "label": "Google Gemma 4 E2B config", "url": "https://huggingface.co/google/gemma-4-E2B/raw/63db66a33dc06d58c02b1e887446e103c202602c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 text/vision/audio configs, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E2B IT config comparison", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible" ], "notes": "Manual comparison found matching text_config, audio_config, vision_config, top-level dtype, model_type, and tie_word_embeddings between the base and instruction-tuned configs. The instruction-tuned config only adds top-level eos_token_id." }, { "label": "Google Gemma 4 E2B Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "total_params_b", "downloads" ], "notes": "The API response at commit 63db66a33dc06d58c02b1e887446e103c202602c records safetensors parameters BF16: 5123178051 and total: 5123178051, any-to-any pipeline, Apache-2.0 licensing, region:us, and 184194 downloads. The direct safetensors header sum is 928 parameters higher, so this profile uses the header count for exact memory bytes." }, { "label": "Google Gemma 4 E2B safetensors header", "url": "https://huggingface.co/google/gemma-4-E2B/resolve/63db66a33dc06d58c02b1e887446e103c202602c/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. The file Content-Length is 10246621918 bytes, with a 263952-byte header and 10246357958 tensor bytes. Stored tensors sum to 5123178979 BF16 parameters / 10.246357958 GB. model.language_model tensors sum to 4647449891 BF16 parameters / 9.294899782 GB. model.language_model.embed_tokens.weight has shape [262144, 1536] and contributes 0.805306368 GB. model.language_model.embed_tokens_per_layer.weight has shape [262144, 8960] and contributes 4.69762048 GB. The header has no separate lm_head.weight, so the standard embedding remains swept as the tied output projection. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens_per_layer, sum to 2298639651 BF16 parameters / 4.597279302 GB. Resident-only tensors, defined as audio tower/projection plus vision tower/projection plus the per-layer embedding table, sum to 2824539328 BF16 parameters / 5.649078656 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served base config, instruction-tuned sibling config comparison, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds on the base Gemma 4 E2B BF16 checkpoint. It deliberately separates resident multimodal and PLE weights from per-token swept language weights." }, { "id": "google--gemma-4-e4b-it-assistant", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/gemma-4-E4B-it-assistant", "title": "Google Gemma 4 E4B IT Assistant Drafter BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16/I64 Gemma 4 E4B IT MTP assistant drafter repo.", "model_family": "gemma4-assistant-speculative-drafter", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "derived_package", "source": "Model card speculative-decoding instructions and generation_config assistant metadata", "config_compatible": false, "notes": "The model card identifies this repo as the Multi-Token Prediction drafter for google/gemma-4-E4B-it. It is loaded as assistant_model alongside the target E4B model in Transformers speculative decoding, not as a replacement target model." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b-assistant-drafter", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.078779908, "swept_params_b": 0.078779908, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.15913268, "swept_weight_gb": 0.15913268, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16_i64", "swept_parameter_scope": "Exact BF16/I64 tensor bytes are recorded, but Bounds Engine v1 does not use them for production throughput because this repo is a speculative drafter paired with a target model.", "auxiliary_scope": "No separate auxiliary resident tensors were identified in the single safetensors file; the repo stores masked_embedding, model.embed_tokens, model.layers, model.norm, pre_projection, and post_projection tensors.", "notes": "The single safetensors file stores 50 tensors totaling 0.159132680 GB: BF16 0.157035528 GB and I64 0.002097152 GB. Tensor groups are model.embed_tokens 0.134217728 GB, model.layers 0.017836552 GB, model.norm 0.000000512 GB, pre_projection.weight 0.002621440 GB, post_projection.weight 0.001310720 GB, and masked_embedding 0.003145728 GB." }, "kv_adapter": { "kind": "unknown", "reason": "Gemma4AssistantForCausalLM is an MTP speculative drafter. Production throughput depends on the paired target model, draft length, acceptance rate, verification schedule, and extra target-side validation work. Bounds Engine v1 only models single-model ordinary autoregressive text-token decode.", "notes": "The config records four assistant text layers with three sliding-attention layers and one full-attention layer, but using those fields as ordinary target-model decode would be misleading because generation_config marks is_assistant true and num_assistant_tokens 6." }, "notes": "This profile intentionally fails closed even though exact weights and assistant-layer geometry are accessible, because comparison tok/s should not treat this drafter as a standalone target model." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2.0199652936888426, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-gemma4-mtp-assistant-drafter", "dequantization_notes": "No quantized weight representation is assumed for this BF16/I64 repo.", "notes": "The repo is used as assistant_model in speculative decoding with google/gemma-4-E4B-it. A production profile needs a paired target+assistant speculative-decoding adapter, not a standalone ordinary text-decode adapter." }, "evidence": [ { "label": "Google Gemma 4 E4B IT Assistant API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-assistant", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 1754649e3d145ddb62afd52e7204d5a755653aec, the API reports an Apache-2.0 any-to-any repo with transformers, safetensors, gemma4_assistant, text-generation, endpoints_compatible, region:us, 334689 downloads, and safetensors metadata for I64 262144 plus BF16 78517764 parameters." }, { "label": "Google Gemma 4 E4B IT Assistant config", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/raw/1754649e3d145ddb62afd52e7204d5a755653aec/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "assistant_geometry", "unsupported_reason" ], "notes": "The config records Gemma4AssistantForCausalLM, model_type gemma4_assistant, BF16 dtype, backbone_hidden_size 2560, 2048 centroids, centroid_intermediate_top_k 32, use_ordered_embeddings true, tie_word_embeddings true, and a four-layer text_config with three sliding-attention layers, one full-attention layer, hidden_size 256, intermediate_size 2048, 4 attention heads, 2 KV heads, attention_k_eq_v false, 512-token sliding window, and 131072 max position embeddings." }, { "label": "Google Gemma 4 E4B IT Assistant generation config", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/raw/1754649e3d145ddb62afd52e7204d5a755653aec/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config explicitly records is_assistant true, num_assistant_tokens 6, and a constant assistant-token schedule." }, { "label": "Google Gemma 4 E4B IT Assistant model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm", "base_model_proof" ], "notes": "The model card states that this checkpoint contains Multi-Token Prediction drafters for Gemma 4 models. Its examples load google/gemma-4-E4B-it as TARGET_MODEL_ID and this repo as ASSISTANT_MODEL_ID, then pass assistant_model=assistant_model to target_model.generate." }, { "label": "Google Gemma 4 E4B IT Assistant safetensors header", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/resolve/1754649e3d145ddb62afd52e7204d5a755653aec/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "resident_weight_gb", "weight_format" ], "notes": "A range-read of the single safetensors header found a 5520-byte header and 50 tensors totaling 0.159132680 GB: BF16 0.157035528 GB and I64 0.002097152 GB. Tensor groups are masked_embedding 0.003145728 GB, model.embed_tokens 0.134217728 GB, model.layers 0.017836552 GB, model.norm 0.000000512 GB, pre_projection 0.002621440 GB, and post_projection 0.001310720 GB. The linked file size is 159138208 bytes, so non-tensor safetensors overhead is 5528 bytes." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from live HF API metadata, model card, served config, generation config, and direct safetensors header byte grouping. Marked unsupported because Bounds Engine v1 lacks a paired target+assistant speculative decoding adapter." }, "unsupported_reason": "Gemma 4 assistant checkpoints are speculative drafters used with a separate target model. Bounds Engine v1 does not model paired target+assistant speculative decoding, acceptance rates, or target verification work, so standalone tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports paired speculative-decoding profiles." }, { "id": "google--gemma-4-e4b-it-qat-q4-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E4B-it-qat-q4_0-gguf", "title": "Google Gemma 4 E4B IT QAT Q4_0 GGUF", "summary": "Audited memory-side text-decode bounds profile for Google's official QAT Q4_0 GGUF artifact of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized base config, Google QAT release notes in the model card, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo metadata identifies this package as a GGUF quantization of google/gemma-4-E4B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 E4B text geometry as the pinned QAT unquantized config. The QAT unquantized model card states that GGUF Q4_0 packages are ready-to-deploy artifacts for broad ecosystem compatibility." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.463013674, "swept_params_b": 4.644441386, "auxiliary_resident_params_b": 2.818572288, "resident_weight_gb": 5.154939136, "swept_weight_gb": 2.82700832, "auxiliary_resident_weight_gb": 2.327930816, "resident_parameter_scope": "selected main GGUF linked file size and header tensor parameters for gemma-4-E4B_q4_0-it.gguf", "swept_parameter_scope": "ordinary text decode charges GGUF tensor spans in the selected main artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; gemma-4-E4B-it-mmproj.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "Gemma 4 E4B uses per-layer embeddings. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per-layer projection tensors inside blk.*, normal language blocks, token_embd.weight, output_norm.weight, per_layer_proj_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The QAT base config records global_head_dim 512 and attention_k_eq_v false, and the selected GGUF header contains separate attn_k and attn_v tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the QAT base config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6907315678596464, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is Google's official main QAT Q4_0 GGUF file because HF API gguf.totalFileSize matches gemma-4-E4B_q4_0-it.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Google Gemma 4 E4B IT QAT Q4_0 GGUF API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-qat-q4_0-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit bb3b92e6f031fa438b409f898dd9f14f499a0cb0 records base_model google/gemma-4-E4B-it-qat-q4_0-unquantized, Apache-2.0 license, any-to-any pipeline, endpoints_compatible, region:us, conversational, 301981 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 7463013674, and gguf.totalFileSize 5154939136." }, { "label": "Google Gemma 4 E4B IT QAT Q4_0 GGUF model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf/raw/bb3b92e6f031fa438b409f898dd9f14f499a0cb0/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card metadata records Apache-2.0 licensing and base_model google/gemma-4-E4B-it-qat-q4_0-unquantized for this official GGUF package." }, { "label": "Google Gemma 4 E4B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable QAT unquantized config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 QAT release model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "qat_release_scope", "gguf_format" ], "notes": "The QAT model card states this release contains quantization-aware trained checkpoints, including unquantized QAT checkpoints and ready-to-deploy GGUF Q4_0 packages, and lists Gemma 4 E4B as a dense 42-layer model with 512-token sliding windows, 128K context, image/audio support, and per-layer embeddings." }, { "label": "Google Gemma 4 E4B IT QAT Q4_0 GGUF linked-object HEAD checks", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf/tree/bb3b92e6f031fa438b409f898dd9f14f499a0cb0", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E4B_q4_0-it.gguf is 5154939136 bytes, exactly matching API gguf.totalFileSize. The repo also ships gemma-4-E4B-it-mmproj.gguf at 991551904 bytes; that sidecar is not included in this main text artifact profile." }, { "label": "Google Gemma 4 E4B QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf/resolve/bb3b92e6f031fa438b409f898dd9f14f499a0cb0/gemma-4-E4B_q4_0-it.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A range-read of the GGUF v3 header found 43 metadata entries and 666 tensors. The selected file is 5.154939136 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15820736. Tensor spans total 5.139118400 GB across 7463013674 logical elements. per_layer_token_embd.weight is 2.312110080 GB / 2818572288 logical elements and is resident-only for this ordinary text-decode profile. Swept tensor spans excluding that lookup table total 2.827008320 GB / 4644441386 logical elements. token_embd.weight is 0.550502400 GB and no output.weight tensor is stored, so token_embd.weight remains swept as tied output-projection traffic. The header records gemma4.block_count 42, context_length 131072, attention.head_count 8, attention.head_count_kv 2, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, embedding_length_per_layer_input 256, and no mmproj, vision, or audio tensors in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned official Google GGUF model card, immutable QAT unquantized config, QAT release model card, HEAD checks for linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for the official Google main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "google--gemma-4-e4b-it-qat-w4a16-ct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E4B-it-qat-w4a16-ct", "title": "Google Gemma 4 E4B IT QAT W4A16 CT", "summary": "Audited memory-side text-decode bounds profile for Google's compressed-tensors W4A16 QAT Gemma 4 E4B IT repo.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API metadata, served compressed-tensors config, Google QAT unquantized config, existing Gemma 4 E4B IT profile comparison, Transformers Gemma 4 implementation, and direct safetensors header metadata", "config_compatible": true, "notes": "The compressed-tensors repo records google/gemma-4-E4B-it-qat-q4_0-unquantized as its base model. Manual comparison found the compressed, QAT unquantized, and audited Gemma 4 E4B IT configs match on the checked text, multimodal, context, attention, PLE, and serving-relevant fields; the compressed repo adds quantization_config." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.736340888, "swept_params_b": 4.768591576, "auxiliary_resident_params_b": 3.967749312, "resident_weight_gb": 11.513497412, "swept_weight_gb": 3.577998788, "auxiliary_resident_weight_gb": 7.935498624, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "model.language_model excluding embed_tokens and embed_tokens_per_layer plus separate lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens input table, model.language_model.embed_tokens_per_layer PLE lookup table, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision tensors are resident for the multimodal PLE package but not full-matrix swept for each generated text token", "notes": "Compressed-tensors stores packed I32 int4 weights, BF16 scale and ignored-module tensors, and tiny I64 weight_shape side tensors. The config records tie_word_embeddings true, but this compressed artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection while treating model.language_model.embed_tokens.weight and model.language_model.embed_tokens_per_layer.weight as resident lookup/input tables for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 4, "read_layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. With num_kv_shared_layers 18, only layers 5, 11, 17, and 23 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 35, "alloc_layers": 20, "read_layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 18, only the first twenty sliding-attention layers allocate K/V cache, while all 35 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E4B shares K/V across the final 18 decoder layers, so allocation layer counts differ from read layer counts. Quantizing weights does not change the BF16 KV cache assumption because the compressed-tensors config has kv_cache_scheme null." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary cached text decode after any multimodal prefill, not encoder or input-projection throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.3178855495227555, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-ple-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 scale tensors, ignored BF16 modules, BF16 embeddings/towers/lm_head, and I64 shape side tensors from safetensors headers. Dequantization, activation traffic, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 32, bfloat16 model dtype, and kv_cache_scheme null. weight_bytes_per_param records the exact resident stored-byte average across header-logical safetensors parameters; resident and swept traffic use the exact header byte fields above." }, "evidence": [ { "label": "Google Gemma 4 E4B IT QAT W4A16 CT API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-qat-w4a16-ct", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The live HF API response at commit ef0a4c43726bde42a3ca04fd300397c0b8b3c3f0 records a public non-gated Apache-2.0 any-to-any repo with transformers, safetensors, compressed-tensors, endpoints_compatible, region:us, base_model google/gemma-4-E4B-it-qat-q4_0-unquantized, 291658 downloads, and API safetensors logical parameters I64: 686, I32: 3972792320, BF16: 4763546954, total: 8736339960. The direct header has 928 additional BF16 elements, so this profile uses direct header counts for exact memory bytes." }, { "label": "Google Gemma 4 QAT model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family", "per_layer_embeddings", "max_context_tokens", "weight_format" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E4B-it-qat-q4_0-unquantized, QAT release packaging, compressed-tensors W4A16 artifacts for native optimized inference with vLLM, any-to-any multimodal support, 128K context for E4B, 42 E4B layers, 512-token sliding window, 262K vocabulary, text/image/audio modalities, about 150M vision parameters, about 300M audio parameters, and Per-Layer Embeddings for E2B/E4B." }, { "label": "Google Gemma 4 E4B IT QAT W4A16 CT config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct/raw/ef0a4c43726bde42a3ca04fd300397c0b8b3c3f0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings", "kv_sharing" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, compressed-tensors pack-quantized int4 weights, group_size 32, symmetric group strategy, quantization_status compressed, kv_cache_scheme null, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 18, two KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config. The explicit quantization ignore list includes lm_head, audio/embed/vision projection tensors, and vision/audio tower modules." }, { "label": "Google Gemma 4 E4B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter", "per_layer_embeddings" ], "notes": "Manual comparison found the QAT unquantized config and this compressed-tensors config match the checked top-level text, vision, audio, dtype, context, attention, PLE, and serving-relevant fields. The same fields also match the already audited Gemma 4 E4B IT config after excluding quantization metadata." }, { "label": "Transformers Gemma 4 implementation and PLE documentation", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py", "source_type": "manual_review", "supports": [ "kv_sharing", "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The upstream Gemma 4 implementation computes first_kv_shared_layer_idx = num_hidden_layers - num_kv_shared_layers, omits k_proj/v_proj/k_norm/v_norm for shared layers, updates past_key_values only for non-shared layers, and still passes shared key/value states to every shared attention layer. This supports charging fewer allocation layers than read layers. The Gemma 4 PLE path uses token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection, supporting resident-only treatment of the large per-layer embedding table while charging projection and block tensors as swept traffic." }, { "label": "Google Gemma 4 E4B IT QAT W4A16 CT safetensors header", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct/resolve/ef0a4c43726bde42a3ca04fd300397c0b8b3c3f0/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings", "kv_sharing" ], "notes": "The single safetensors file was range-read directly. The linked Content-Length is 11513861828 bytes, with a 364408-byte header and 11513497412 tensor bytes across 2763 tensors. Stored tensors sum to 8.736340888B header-logical parameters / 11.513497412 GB: 343 I32 packed tensors carrying 3.972792320B logical int4 weights / 1.986396160 GB, 2077 BF16 tensors carrying 4.763547882B parameters / 9.527095764 GB, and 343 I64 shape tensors carrying 686 values / 0.000005488 GB. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens and embed_tokens_per_layer plus lm_head, sum to 4.768591576B logical parameters / 3.577998788 GB. Resident-only tensors, defined as model.language_model.embed_tokens, model.language_model.embed_tokens_per_layer, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision tensors, sum to 3.967749312B parameters / 7.935498624 GB. Header grouping confirms k_proj/v_proj tensors are absent from layers 24-41, matching num_kv_shared_layers 18." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable served config, QAT unquantized config comparison, existing Gemma 4 E4B profile comparison, upstream Transformers Gemma 4 implementation review, and direct safetensors header byte grouping." }, "notes": "Use this profile for Google's official W4A16 compressed-tensors artifact. It models ordinary cached text decode after any multimodal prefill and deliberately separates resident input/PLE/multimodal tensors from per-token swept language/logit weights." }, { "id": "google--gemma-4-e4b-it", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E4B-it", "title": "Google Gemma 4 E4B IT BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 E4B instruction-tuned repo.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B", "relation": "finetune", "source": "Hugging Face model card base_model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text, vision, audio, dtype, context, and attention fields between the instruction-tuned and base configs. The instruction-tuned config adds a top-level eos_token_id list, which does not change the memory-side architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 15.992314836, "swept_weight_gb": 9.398993492, "auxiliary_resident_weight_gb": 6.593321344, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers_excluding_per_layer_embedding_table", "auxiliary_scope": "model.audio_tower, model.embed_audio, model.vision_tower, model.embed_vision, and model.language_model.embed_tokens_per_layer.weight are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "Gemma 4 E4B uses per-layer embeddings. The large model.language_model.embed_tokens_per_layer.weight table is resident and looked up during text decode, not read as a full matrix per token. The standard model.language_model.embed_tokens.weight table remains in swept traffic because it also serves the output projection in this tied-embedding repo." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and tensor headers contain separate k_proj and v_proj weights." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level, text, vision, and audio configs record bfloat16 dtype, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 E4B IT model card", "url": "https://huggingface.co/google/gemma-4-E4B-it", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "ple_resident_split" ], "notes": "The model metadata identifies google/gemma-4-E4B as the base model, Apache-2.0 licensing, and any-to-any multimodal packaging. The card states E4B has 4.5B effective parameters, 8B with embeddings, 42 layers, 128K context, about 150M vision encoder parameters, about 300M audio encoder parameters, and per-layer embeddings." }, { "label": "Google Gemma 4 E4B IT config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 text/vision/audio configs, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E4B base config", "url": "https://huggingface.co/google/gemma-4-E4B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text, multimodal, serving, and context architecture fields match the instruction-tuned repo config." }, { "label": "Google Gemma 4 E4B IT Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "base_model_proof", "total_params_b" ], "notes": "The API response at commit fee6332c1abaafb77f6f9624236c63aa2f1d0187 records safetensors parameters BF16: 7996156490 and total: 7996156490, any-to-any pipeline, and base_model google/gemma-4-E4B. The direct safetensors header sum is 928 parameters higher, so this profile uses the header count for exact memory bytes." }, { "label": "Google Gemma 4 E4B IT safetensors header", "url": "https://huggingface.co/google/gemma-4-E4B-it/resolve/main/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. Stored tensors sum to 7996157418 BF16 parameters and 15.992314836 GB. model.language_model.embed_tokens.weight is 671088640 BF16 params and 1.34217728 GB. model.language_model.embed_tokens_per_layer.weight is 2818572288 BF16 params and 5.637144576 GB. The header has no separate lm_head.weight. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens_per_layer.weight but including the tied standard embedding/output projection, sum to 4699496746 BF16 parameters and 9.398993492 GB. Resident-only tensors, defined as audio tower/projection plus vision tower/projection plus the per-layer embedding table, sum to 3296660672 BF16 parameters and 6.593321344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from served config, base config comparison, model card/API metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal and PLE weights from per-token swept language weights." }, { "id": "google--gemma-4-e4b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "google/gemma-4-E4B", "title": "Google Gemma 4 E4B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Gemma 4 E4B pretrained repo.", "model_family": "gemma4-dense-multimodal-ple", "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 15.992314836, "swept_weight_gb": 9.398993492, "auxiliary_resident_weight_gb": 6.593321344, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model_safetensors_headers_excluding_per_layer_embedding_table", "auxiliary_scope": "model.audio_tower, model.embed_audio, model.vision_tower, model.embed_vision, and model.language_model.embed_tokens_per_layer.weight are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "Gemma 4 E4B uses per-layer embeddings. The large model.language_model.embed_tokens_per_layer.weight table is resident and looked up during text decode, not read as a full matrix per token. The standard model.language_model.embed_tokens.weight table remains in swept traffic because it also serves the output projection in this tied-embedding repo." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and tensor headers contain separate K and V projection weights." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The top-level, text, vision, and audio configs record bfloat16 dtype, and the safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Google Gemma 4 E4B model card", "url": "https://huggingface.co/google/gemma-4-E4B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "architecture", "ple_resident_split" ], "notes": "The card identifies this as an Apache-2.0 any-to-any Gemma 4 base model. It states E4B has 4.5B effective parameters, 8B with embeddings, 42 layers, 512-token sliding windows, 128K context, 262K vocabulary, image and audio support, about 150M vision encoder parameters, about 300M audio encoder parameters, and per-layer embeddings." }, { "label": "Google Gemma 4 E4B config", "url": "https://huggingface.co/google/gemma-4-E4B/raw/a24c9379fd3839ae84e97f0b6aa3152fce9bd033/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, bfloat16 top-level/text/vision/audio configs, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E4B IT config comparison", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "architecture" ], "notes": "Manual comparison against the audited instruction-tuned Gemma 4 E4B profile found matching top architecture, model type, dtype, tied embedding setting, text config, vision config, audio config, and vision soft-token fields. The instruction-tuned config only adds a top-level eos_token_id list, which does not change memory-side architecture." }, { "label": "Google Gemma 4 E4B Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B", "source_type": "derived_calculation", "supports": [ "repo", "weight_format", "downloads", "total_params_b" ], "notes": "The API response at commit a24c9379fd3839ae84e97f0b6aa3152fce9bd033 records safetensors parameters BF16: 7996156490 and total: 7996156490, any-to-any pipeline, Apache-2.0 license, region:us, and 532775 downloads. The direct safetensors header sum is 928 parameters higher, so this profile uses the header count for exact memory bytes." }, { "label": "Google Gemma 4 E4B safetensors header", "url": "https://huggingface.co/google/gemma-4-E4B/resolve/a24c9379fd3839ae84e97f0b6aa3152fce9bd033/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file header was range-read directly. Stored tensors sum to 7996157418 BF16 parameters and 15.992314836 GB. model.language_model.embed_tokens.weight is 671088640 BF16 params and 1.342177280 GB. model.language_model.embed_tokens_per_layer.weight is 2818572288 BF16 params and 5.637144576 GB. The header has no separate lm_head.weight. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens_per_layer.weight but including the tied standard embedding/output projection, sum to 4699496746 BF16 parameters and 9.398993492 GB. Resident-only tensors, defined as audio tower/projection plus vision tower/projection plus the per-layer embedding table, sum to 3296660672 BF16 parameters and 6.593321344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, instruction-tuned profile/config comparison, linked-object HEAD check, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds on the pretrained Gemma 4 E4B package. It deliberately separates resident multimodal and PLE weights from per-token swept language weights." }, { "id": "google--medgemma-1-5-4b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/medgemma-1.5-4b-it", "title": "Google MedGemma 1.5 4B IT BF16", "summary": "Unsupported profile stub for the auto-gated BF16 MedGemma 1.5 4B instruction-tuned repo.", "model_family": "medgemma-gemma3", "architecture": { "canonical_architecture_id": "medgemma-1.5-4b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 4.300079472, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 4300079472 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, generation config, safetensors index, safetensors shards, ranged safetensors bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer MedGemma/Gemma 3 layer count, KV heads, context length, sliding-window behavior, vision/text split, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google MedGemma 1.5 4B IT API metadata", "url": "https://huggingface.co/api/models/google/medgemma-1.5-4b-it", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "serving", "unsupported_reason" ], "notes": "At repo SHA 91850547d9f0b2fdd21aa7c5f4f3d1a8a52c243b, the API reports gated:auto, image-text-to-text pipeline, transformers library, Health AI Developer Foundations terms, region:us, medical/radiology/clinical tags, and BF16 safetensors count 4300079472. Current downloads are 310950." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/medgemma-1.5-4b-it/raw/91850547d9f0b2fdd21aa7c5f4f3d1a8a52c243b/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, generation config, and safetensors index requests returned 401 restricted-repo responses. The local HF CLI is authenticated as osolmaz, but hf download config.json returned 'Access denied. This repository requires approval.' HEAD and ranged requests for both safetensors shards and model.safetensors.index.json also returned 401." } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, vision/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--medgemma-4b-it", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/medgemma-4b-it", "title": "Google MedGemma 4B IT BF16", "summary": "Unsupported profile stub for the auto-gated BF16 MedGemma 4B instruction-tuned repo.", "model_family": "medgemma-gemma3", "architecture": { "canonical_architecture_id": "medgemma-4b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 4.300079472, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 4300079472 BF16 safetensors parameters for this repo. KV geometry and swept traffic are not audited because the config and safetensors headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, safetensors index, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer MedGemma/Gemma 3 layer count, KV heads, context length, sliding-window behavior, vision/text split, tied embeddings, or swept decode traffic from the model name or API config summary. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "The API config summary identifies Gemma3ForConditionalGeneration, but the raw config and tensor headers are inaccessible. This profile intentionally fails closed until the gated files can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google MedGemma 4B IT API metadata", "url": "https://huggingface.co/api/models/google/medgemma-4b-it", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "serving", "unsupported_reason" ], "notes": "At repo SHA 290cda5eeccbee130f987c4ad74a59ae6f196408, the API reports gated:auto, image-text-to-text pipeline, transformers library, Health AI Developer Foundations terms, base_model google/medgemma-4b-pt, region:us, medical/radiology/clinical tags, and BF16 safetensors count 4300079472. Current downloads are 291353." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/medgemma-4b-it/raw/290cda5eeccbee130f987c4ad74a59ae6f196408/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and safetensors index requests returned 401 restricted-repo responses. The local HF CLI is authenticated as osolmaz, but hf download config.json, model.safetensors.index.json, and README.md returned 'Access denied. This repository requires approval.'" } ], "unsupported_reason": "Gated config and tensor headers are not accessible in this audit environment, so KV geometry, max context, vision/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--paligemma-3b-ft-cococap-448", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/paligemma-3b-ft-cococap-448", "title": "Google PaliGemma 3B FT COCOCap 448 F32", "summary": "Unsupported profile stub for the manually gated F32 PaliGemma 3B COCO captioning fine-tune.", "model_family": "paligemma-multimodal", "architecture": { "canonical_architecture_id": "paligemma-3b-ft-cococap-448", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.924351216, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2924351216 F32 safetensors parameters for this repo. KV geometry, text/vision tensor split, tied embedding behavior, and swept ordinary text traffic are not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw README, config, safetensors index, safetensors shard bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer PaliGemma text decoder layer count, KV heads, head dimension, context length, image-token behavior, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "F32 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google PaliGemma 3B FT COCOCap 448 API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-ft-cococap-448", "source_type": "model_card", "supports": [ "repo", "pipeline", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit cb5ff322983f2ebcd916c3394ffcef78ed4e9a9b, the API reports gated: manual, Transformers image-text-to-text pipeline, Gemma license, PaliGemma tags, region:us, current downloads 231,962, and safetensors parameters F32 2,924,351,216. The API exposes only a partial top-level config with PaliGemmaForConditionalGeneration and tokenizer metadata, not the text/vision geometry needed for bounds." }, { "label": "Gated README, config, and safetensors access checks", "url": "https://huggingface.co/google/paligemma-3b-ft-cococap-448/raw/cb5ff322983f2ebcd916c3394ffcef78ed4e9a9b/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Pinned raw README, config, model.safetensors.index.json, and a ranged model-00001-of-00003.safetensors request returned 401 restricted access. hf download config.json with the configured CLI identity returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context, multimodal/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--paligemma-3b-mix-224", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/paligemma-3b-mix-224", "title": "Google PaliGemma 3B Mix 224 F32", "summary": "Unsupported profile stub for the manually gated F32 PaliGemma 3B mix multimodal repo.", "model_family": "paligemma-multimodal", "architecture": { "canonical_architecture_id": "paligemma-3b-mix-224", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.92346648, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2923466480 F32 safetensors parameters for this repo. KV geometry, text/vision tensor split, tied embedding behavior, and swept ordinary text traffic are not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, safetensors index, safetensors shard bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer PaliGemma text decoder layer count, KV heads, head dimension, context length, image-token behavior, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "F32 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google PaliGemma 3B Mix 224 API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-mix-224", "source_type": "model_card", "supports": [ "repo", "pipeline", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit d1d8734c9c3ad0ccfeea4afc270faa356c2ba515, the API reports gated: manual, Transformers image-text-to-text pipeline, Gemma license, PaliGemma tags, region:us, current downloads 227287, and safetensors parameters F32 2923466480. The API exposes only a partial top-level config with PaliGemmaForConditionalGeneration and tokenizer metadata, not the text/vision geometry needed for bounds." }, { "label": "Gated README, config, and safetensors access checks", "url": "https://huggingface.co/google/paligemma-3b-mix-224/raw/d1d8734c9c3ad0ccfeea4afc270faa356c2ba515/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Pinned raw README, config, model.safetensors.index.json, and a ranged model-00001-of-00003.safetensors request returned 401 restricted access. hf download config.json with the configured CLI identity osolmaz returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context, multimodal/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--paligemma-3b-pt-224", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/paligemma-3b-pt-224", "title": "Google PaliGemma 3B PT 224 F32", "summary": "Unsupported profile stub for the manually gated F32 PaliGemma 3B pretraining multimodal repo.", "model_family": "paligemma-multimodal", "architecture": { "canonical_architecture_id": "paligemma-3b-pt-224", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.92346648, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2923466480 F32 safetensors parameters for this repo. KV geometry, text/vision tensor split, tied embedding behavior, and swept ordinary text traffic are not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, safetensors index, safetensors shard bytes, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer PaliGemma text decoder layer count, KV heads, head dimension, context length, image-token behavior, tied embeddings, or swept decode traffic from the model name or scraped metadata. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor scope are unavailable.", "notes": "F32 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Google PaliGemma 3B PT 224 API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-pt-224", "source_type": "model_card", "supports": [ "repo", "pipeline", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 35e4f46485b4d07967e7e9935bc3786aad50687c, the API reports gated: manual, Transformers image-text-to-text pipeline, Gemma license, PaliGemma tags, region:us, current downloads 278336, and safetensors parameters F32 2923466480. The API exposes only a partial top-level config with PaliGemmaForConditionalGeneration and tokenizer metadata, not the text/vision geometry needed for bounds." }, { "label": "Google PaliGemma 3B PT 224 model card frontmatter", "url": "https://huggingface.co/google/paligemma-3b-pt-224/raw/35e4f46485b4d07967e7e9935bc3786aad50687c/README.md", "source_type": "model_card", "supports": [ "license", "pipeline", "unsupported_reason" ], "notes": "The public README frontmatter records library_name transformers, license gemma, image-text-to-text pipeline, and a manual gated access prompt requiring license acknowledgement. It does not expose executable architecture, KV, or tensor layout details." }, { "label": "Gated config and safetensors access checks", "url": "https://huggingface.co/google/paligemma-3b-pt-224/raw/35e4f46485b4d07967e7e9935bc3786aad50687c/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Pinned raw config returned 401 restricted access. Pinned model.safetensors.index.json and ranged model-00001-of-00003.safetensors requests returned 401 restricted access. hf download config.json with the configured CLI identity returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context, multimodal/text tensor split, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "google--reformer-crime-and-punishment", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "google/reformer-crime-and-punishment", "title": "Google Reformer Crime and Punishment FP32", "summary": "Unsupported profile stub with exact FP32 checkpoint tensor evidence for the small Reformer language model.", "model_family": "reformer-local-lsh-decoder", "architecture": { "canonical_architecture_id": "reformer-crime-and-punishment", "max_context_tokens": 524288, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.002748544, "swept_params_b": 0.002437248, "auxiliary_resident_params_b": 0.000311296, "resident_weight_gb": 0.010994176, "swept_weight_gb": 0.009748992, "auxiliary_resident_weight_gb": 0.001245184, "resident_parameter_scope": "pinned pytorch_model.bin FP32 state_dict tensor storages", "swept_parameter_scope": "all non-embedding FP32 tensors are recorded as naive AR swept traffic evidence only because production bounds are disabled", "auxiliary_scope": "word embeddings and axial position embeddings are resident; per-token lookup traffic is not represented as full-matrix swept traffic", "notes": "A torch-free metadata unpickler read the pinned legacy PyTorch pickle state_dict and found 77 FloatStorage tensors totaling 2,748,544 FP32 parameters / 10,994,176 bytes. Embedding and axial-position tensors total 311,296 parameters / 1,245,184 bytes. Non-embedding tensors, including attention projections, feed-forward weights, layer norms, and lm_head decoder/bias tensors, total 2,437,248 parameters / 9,748,992 bytes." }, "kv_adapter": { "kind": "unknown", "reason": "Reformer autoregressive generation does not use the normal per-layer K/V cache modeled by Bounds Engine v1. The Transformers implementation uses ReformerDynamicCache with per-layer cached hidden states plus LSH bucket assignments, and the config alternates local and LSH attention layers with chunking and hashing.", "notes": "The config records local/lsh/local/lsh/local/lsh layers, 64-token local and LSH chunks, one hash, buckets [64, 128], one previous chunk, and output_past true. A production adapter needs Reformer-specific state/cache traffic for hidden-state cache growth, LSH bucket storage, hashing/gather operations, local chunk windows, and the exact incremental decode path." }, "notes": "This profile intentionally fails closed. Treating the model as dense full-context KV would publish misleading tok/s numbers because the cache is past_buckets_states, not standard K/V tensors." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 4, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 4, "runtime_format": "unsupported-reformer-local-lsh-transformers", "dequantization_notes": "No quantized weight representation is assumed for this FP32 PyTorch checkpoint.", "notes": "The checkpoint is a legacy PyTorch pickle state_dict plus a Rust .ot sibling. Bounds Engine v1 disables production throughput until it has a Reformer cache/state adapter." }, "evidence": [ { "label": "Google Reformer Crime and Punishment API metadata", "url": "https://huggingface.co/api/models/google/reformer-crime-and-punishment", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "pipeline", "tags", "weight_artifacts" ], "notes": "At commit 8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac, the live API records a public non-gated text-generation Transformers repo with Reformer tags, region:us, current downloads 105506, siblings including pytorch_model.bin and rust_model.ot, and no explicit license tag." }, { "label": "Google Reformer Crime and Punishment config", "url": "https://huggingface.co/google/reformer-crime-and-punishment/raw/8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "layers", "attention_heads", "head_dim", "unsupported_reason" ], "notes": "The pinned config records ReformerModelWithLMHead, model_type reformer, is_decoder true, output_past true, 6 hidden layers with attn_layers [local, lsh, local, lsh, local, lsh], hidden size 256, 2 attention heads, 64 attention head size, local and LSH chunk length 64, num_hashes 1, num_buckets [64, 128], num_chunks_before 1, num_chunks_after 0, axial position embeddings with shape [512, 1024], and 524288 max position embeddings." }, { "label": "Google Reformer Crime and Punishment model card", "url": "https://huggingface.co/google/reformer-crime-and-punishment/raw/8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac/README.md", "source_type": "model_card", "supports": [ "training_provenance", "runtime_format" ], "notes": "The pinned card says this is a Reformer model trained on the Crime and Punishment text, originally trained in Flax using the Google Trax Reformer text-generation notebook, then converted to Hugging Face's PyTorch ReformerLM model ReformerModelWithLMHead." }, { "label": "Google Reformer Crime and Punishment PyTorch checkpoint metadata", "url": "https://huggingface.co/google/reformer-crime-and-punishment/resolve/8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac/pytorch_model.bin", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format" ], "notes": "Linked-object HEAD reports pytorch_model.bin is 11013576 bytes. A local metadata-only unpickle of the legacy PyTorch state_dict, with fake torch storage classes and no tensor materialization, found 77 FloatStorage tensors totaling 10994176 tensor bytes / 2748544 FP32 parameters. Tensor byte groups are embeddings 1245184, attention projections 1966080, attention output 786432, feed-forward 6309888, layer norms 28672, and lm_head/other 657920. The sibling rust_model.ot linked object is 11014133 bytes." }, { "label": "Transformers Reformer cache implementation", "url": "https://github.com/huggingface/transformers/blob/v4.55.0/src/transformers/models/reformer/modeling_reformer.py", "source_type": "manual_review", "supports": [ "unsupported_reason", "kv_adapter" ], "notes": "Manual review of the Transformers Reformer implementation found ReformerDynamicCache stores states_cache and buckets_cache and converts to/from legacy past_buckets_states. LSHSelfAttention accepts past_buckets_states, hashes query-key vectors into buckets, gathers sorted chunks, and updates cached hidden states and bucket assignments. This is not the standard K/V cache traffic modeled by Bounds Engine v1." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from live HF API metadata, pinned config, pinned model card, linked-object HEAD checks, metadata-only PyTorch state_dict unpickling, and Transformers Reformer runtime source. Marked unsupported because Bounds Engine v1 lacks a Reformer local/LSH cache adapter." }, "unsupported_reason": "Bounds Engine v1 does not model Reformer local/LSH attention cache traffic. The model uses cached hidden states plus LSH bucket assignments rather than ordinary per-layer K/V tensors, so dense full-context or sliding-window KV estimates would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine has a Reformer adapter with hidden-state cache, bucket-cache, hashing, chunking, and local/LSH gather traffic." }, { "id": "groxaxo--qwen3-6-27b-gptq-pro-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit", "title": "groxaxo Qwen3.6 27B GPTQ-Pro 4-bit", "summary": "Audited memory-side text-decode bounds profile for the groxaxo GPTQ-Marlin 4-bit package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The model card and API metadata identify this package as a quantized derivative of Qwen/Qwen3.6-27B. Manual comparison found matching checked top-level, text, and vision architecture fields: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, and 262144 max position embeddings. The package adds GPTQ quantization metadata and omits top-level model.mtp.* tensors from the safetensors index." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 18.709748192, "swept_weight_gb": 15.2454912, "auxiliary_resident_weight_gb": 3.464256992, "resident_parameter_scope": "logical Qwen3.6 27B language-plus-vision parameter split with no model.mtp tensors, plus direct GPTQ safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode; no top-level model.mtp tensors are present in this artifact", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 qweight/qzeros/g_idx tensors, F16 scales, and BF16 embeddings, output head, visual tensors, and selected linear-attention tensors. Logical parameter counts follow the audited Qwen3.6 27B base language/vision split after removing the absent MTP tensor group. Storage metadata tensors are byte traffic, not separate model parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The GPTQ artifact preserves the base Qwen3.6 linear-attention geometry. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with MTP/speculative decoding disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6839176018786363, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-gptq-marlin-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ safetensors bytes: packed I32 weights/zero points/group indexes, F16 scales, and unquantized BF16 embeddings, output head, visual tensors, and selected linear-attention tensors. Dequantization, activation traffic, Marlin kernel behavior, and compute overhead are outside this memory-side bound.", "notes": "The model card recommends vLLM with --dtype float16 and --quantization gptq_marlin. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "groxaxo Qwen3.6 27B GPTQ-Pro API metadata", "url": "https://huggingface.co/api/models/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 62b96ffbca486698719ec67f7f414251ad2347ea, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-27B, with transformers, qwen3_5, gptq, marlin, foem, endpoints_compatible, 4-bit, and region:us tags. Current downloads are 602309. The API safetensors block reports BF16: 3029765360, I32: 24326963200, total: 27356728560 storage-accounting elements." }, { "label": "groxaxo Qwen3.6 27B GPTQ-Pro README", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format", "serving" ], "notes": "The README records base_model Qwen/Qwen3.6-27B, GPTQ-Pro/FOEM quantization, Marlin optimization, vLLM startup with --dtype float16 and --quantization gptq_marlin, and notes that MTP/speculative decoding is optional and not effective in the recorded local test." }, { "label": "groxaxo Qwen3.6 27B GPTQ-Pro config", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit/raw/62b96ffbca486698719ec67f7f414251ad2347ea/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, GPTQ 4-bit quantization with group_size 128, desc_act false, symmetric quantization, pack_dtype int32, lm_head false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and text_config MTP settings." }, { "label": "groxaxo Qwen3.6 27B GPTQ-Pro quantize config", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit/raw/62b96ffbca486698719ec67f7f414251ad2347ea/quantize_config.json", "source_type": "config", "supports": [ "serving", "weight_format" ], "notes": "The quantize_config records method gptq, quant_method gptq, checkpoint_format gptq, format gptq, 4 bits, group_size 128, desc_act false, symmetric quantization, pack_dtype int32, lm_head false, and quantizer gptqmodel:6.1.0-dev." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in checked top-level, text_config, and vision_config geometry fields between the current base BF16 repo and this GPTQ artifact. The artifact adds quantization_config and serving metadata but preserves the architecture fields used by this profile." }, { "label": "groxaxo Qwen3.6 27B GPTQ-Pro safetensors index and shard headers", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit/resolve/62b96ffbca486698719ec67f7f414251ad2347ea/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index maps 2384 tensors across five shards. Its metadata total_size is 18710050584 bytes, but direct range-read safetensors headers sum to 18.709748192 GB, so this profile uses header-derived bytes. Header dtype bytes are BF16 6.059530720 GB, I32 12.270108672 GB, and F16 0.380108800 GB. Stored suffix totals are qweight 12.163481600 GB, qzeros 0.095027200 GB, g_idx 0.011599872 GB, scales 0.380108800 GB, and BF16 weight tensors 6.058829568 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 15.245491200 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual, totals 3.464256992 GB. Header buckets are language_other 12.702694400 GB, lm_head 2.542796800 GB, input embedding 2.542796800 GB, and visual 0.921460192 GB. No model.mtp.* tensors are present in the index." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, README, pinned GPTQ config and quantize_config, base config comparison, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted GPTQ group indexes/scales plus unquantized embeddings, output head, visual tensors, and selected linear-attention tensors." }, { "id": "hcompany--holo-3-1-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Hcompany/Holo-3.1-35B-A3B-GGUF", "title": "Holo 3.1 35B A3B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of Holo 3.1 35B A3B.", "model_family": "holo3.1-qwen3.6-moe-gguf", "base_model_proof": { "base_model": "Hcompany/Holo-3.1-35B-A3B", "relation": "derived_package", "source": "GGUF model card/API metadata, non-GGUF Holo config/API metadata, Qwen base config, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo card documents Holo3.1-35B-A3B and identifies Qwen/Qwen3.6-35B-A3B among the base models. The public non-GGUF Holo config records Qwen3_5MoeForConditionalGeneration with the same memory-relevant text geometry as the selected GGUF header and Qwen base config: 40 text blocks, full_attention_interval 4, 10 full-attention layers, 30 DeltaNet linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 256 routed experts, 8 experts per token, one shared expert, and 262144 context length." }, "architecture": { "canonical_architecture_id": "holo-3-1-35b-a3b-qwen35moe", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 69.376637024, "main_resident_weight_gb": 68.348529152, "auxiliary_resident_weight_gb": 1.028107872, "fixed_weight_gb": 3.924019712, "routed_expert_weight_gb": 0.25165824, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected main.bf16.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected BF16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; q4_k_m.gguf, mmproj.f16.gguf, and imatrix.gguf are sibling sidecars and are not included unless explicitly selected by another workload", "shared_expert_notes": "The Holo/Qwen config records shared_expert_intermediate_size 512, and the selected GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B marketing parameters. The selected BF16 GGUF stores BF16 tensors plus tiny F32 tensors. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The non-GGUF Holo config, Qwen base config, and selected GGUF metadata record 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Holo/Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 main GGUF artifact. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and multimodal prefill require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-moe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected BF16 GGUF artifact. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, image prefill, and write traffic are outside Bounds Engine v1.", "notes": "The API-selected artifact is main.bf16.gguf because HF API gguf.totalFileSize exactly matches that linked object. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param records nominal BF16 tensor payload size." }, "evidence": [ { "label": "Holo 3.1 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/Hcompany/Holo-3.1-35B-A3B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 95dd5447a6f101de041c7f03821dafac2bac19a4 records a public Apache-2.0 image-text-to-text GGUF repo with 191258 downloads, region:us, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 69376637024. The API totalFileSize matches main.bf16.gguf, so this profile targets that artifact." }, { "label": "Holo 3.1 35B A3B GGUF model card", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF/raw/95dd5447a6f101de041c7f03821dafac2bac19a4/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "runtime_format" ], "notes": "The card documents Holo3.1 as a VLM family for computer-use agents, lists Holo3.1-35B-A3B, records base models including Qwen/Qwen3.6-35B-A3B, advertises BF16, FP8, NVFP4, and Q4 GGUF quantizations, and records Apache-2.0 licensing." }, { "label": "Holo 3.1 35B A3B non-GGUF API metadata", "url": "https://huggingface.co/api/models/Hcompany/Holo-3.1-35B-A3B", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family" ], "notes": "The non-GGUF repo is public, Apache-2.0 licensed, image-text-to-text, and records qwen3_5_moe plus multimodal computer-use tags at commit 2bdb92851a8cd9d72cdd891fdf38cfcc7fefae2c. Its API safetensors block reports 35107181936 BF16 parameters, while the pinned config gives the memory-relevant architecture used for compatibility checks. The selected main GGUF artifact excludes separate multimodal sidecars." }, { "label": "Holo 3.1 35B A3B non-GGUF config", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B/raw/2bdb92851a8cd9d72cdd891fdf38cfcc7fefae2c/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, and a 27-layer Qwen3.5 vision config." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "linear_attention_state" ], "notes": "The public Qwen base config records the same audited text decode geometry as Holo and the selected GGUF header, with bfloat16 text dtype. This confirms the Holo text stack preserves the Qwen3.6 35B A3B memory-relevant layout." }, { "label": "Holo 3.1 35B A3B BF16 GGUF linked object and range-read tensor index", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF/resolve/95dd5447a6f101de041c7f03821dafac2bac19a4/main.bf16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 40 metadata entries and 733 tensors. The linked file is 69.376637024 GB. Tensor spans sum to 69.365647872 GB; metadata/tokenizer/header/file overhead accounts for 0.010989152 GB. Tensor spans split into BF16 69.276794880 GB and F32 0.088852992 GB. token_embd.weight is 1.017118720 GB and resident-only; output.weight is a separate 1.017118720 GB swept tensor. Routed expert tensors sum to 64.424509440 GB, or 0.251658240 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 3.924019712 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found main.bf16.gguf 69.376637024 GB, q4_k_m.gguf 21.297870912 GB, mmproj.f16.gguf 0.899282976 GB, and imatrix.gguf 0.192232256 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, non-GGUF Holo config/API metadata, Qwen base config, selected linked file sizes, a direct GGUF header/tensor-index range read of the API-selected BF16 artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected Holo 3.1 35B A3B BF16 main GGUF artifact in ordinary text-decode bounds. Do not infer q4_k_m.gguf, mmproj.f16.gguf, imatrix, multimodal prefill, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "hugging-quants--llama-3-2-1b-instruct-q8-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF", "title": "Hugging Quants Llama 3.2 1B Instruct GGUF Q8_0", "summary": "Audited memory-side text-decode bounds profile for the Q8_0 GGUF artifact of Llama 3.2 1B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-1B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, and GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.2-1B-Instruct as the quantized base. The base raw config is gated in this audit environment, so this profile uses the selected public GGUF header as the direct architecture source instead of copying the base config." }, "architecture": { "canonical_architecture_id": "llama-3.2-1b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.235814432, "swept_params_b": 1.235814432, "resident_weight_gb": 1.3210792, "swept_weight_gb": 1.313251456, "auxiliary_resident_weight_gb": 0.007827744, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for llama-3.2-1b-instruct-q8_0.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected Q8_0 linked file is 1.321079200 GB. Header tensor spans total 1.313251456 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007827744 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 16-layer decoder with 8 KV heads and 64-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected Q8_0 GGUF artifact. It uses the public GGUF metadata for architecture because the Meta base config is gated." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 1.0689947987271928, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q8-0-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact uses Q8_0 tensors plus small F32 norm/rope tensors. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Hugging Quants Llama 3.2 1B Instruct Q8_0 GGUF API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 8790ecfe0c6c7a20f2a532acdfc1b5d98e1bfa8b, the API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.2-1B-Instruct, region:us, 573463 downloads, GGUF architecture llama, 131072 context length, gguf.total 1235814432, and gguf.totalFileSize 1321079200." }, { "label": "Hugging Quants Llama 3.2 1B Instruct Q8_0 GGUF model card", "url": "https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact" ], "notes": "The repo metadata records the GGUF package as a quantized derivative of meta-llama/Llama-3.2-1B-Instruct and exposes a single llama-3.2-1b-instruct-q8_0.gguf artifact." }, { "label": "Llama 3.2 1B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 9213176726f574b556790deb65791e0c5aa438b6, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 1235814400 parameters." }, { "label": "Llama 3.2 1B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A direct raw config request returned 401 restricted access. The profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "Llama 3.2 1B Instruct Q8_0 GGUF linked-object HEAD check", "url": "https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF/tree/8790ecfe0c6c7a20f2a532acdfc1b5d98e1bfa8b", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "The repo contains one GGUF file, llama-3.2-1b-instruct-q8_0.gguf. The selected artifact's HEAD response returned content-length 1321079200 bytes, matching API gguf.totalFileSize." }, { "label": "Llama 3.2 1B Instruct Q8_0 GGUF range-read tensor index", "url": "https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF/resolve/8790ecfe0c6c7a20f2a532acdfc1b5d98e1bfa8b/llama-3.2-1b-instruct-q8_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 30 metadata entries and 147 tensors. The linked file is 1.321079200 GB. Tensor spans sum to 1.313251456 GB: token_embd.weight 0.279085056 GB, blk.* tensors 1.034158080 GB, output_norm.weight 0.000008192 GB, and rope_freqs.weight 0.000000128 GB. Metadata/tokenizer/header/file overhead accounts for 0.007827744 GB. Stored tensor bytes split into Q8_0 1.312980992 GB and F32 0.000270464 GB. The header records llama.block_count 16, context_length 131072, embedding_length 2048, feed_forward_length 8192, attention.head_count 32, attention.head_count_kv 8, attention key/value length 64, rope.freq_base 500000, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, base-model API metadata, gated base-config access check, linked-object HEAD check, and a direct GGUF header/tensor-index range read of the selected Q8_0 artifact." }, "notes": "Use this profile for the selected Llama 3.2 1B Instruct Q8_0 GGUF artifact. Do not infer the gated base config directly; the architecture evidence is the selected GGUF header metadata." }, { "id": "hugging-quants--meta-llama-3-1-70b-instruct-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4", "title": "Meta Llama 3.1 70B Instruct AWQ INT4", "summary": "Audited memory-side bounds profile for the Hugging Quants AWQ INT4 Llama 3.1 70B Instruct repo.", "model_family": "llama3.1-dense-awq", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-70B-Instruct", "relation": "quantized", "source": "Hugging Face model card, served AWQ config, gated base-model API metadata, and safetensors header review", "config_compatible": false, "notes": "The model card states this is a community quantized version of meta-llama/Meta-Llama-3.1-70B-Instruct and provides the AutoAWQ quantization script with model_path set to that gated base. The base API metadata is visible, but raw base config and tensor index access remain gated in this audit environment, so direct config compatibility with the base cannot be independently verified. This profile therefore audits the served AWQ repo directly from its public config and tensor headers." }, "architecture": { "canonical_architecture_id": "llama-3-1-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 39.767785472, "swept_weight_gb": 37.666439168, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "logical Llama 3.1 70B parameters represented by safetensors qweight plus F16 tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors with F16 scales and unquantized F16 embedding/head/norm tensors. Logical parameter counts use the gated base API total and the served tensor layout: I32 qweight tensors are unpacked 8x for logical weight count, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 80 layers, 8 KV heads, hidden size 8192, 64 attention heads, Llama 3 RoPE scaling, and no sliding-window setting, so this profile charges full-context K and V streams for ordinary cached text decode." }, "notes": "Dense LlamaForCausalLM AWQ profile using the served Hugging Quants repo config and tensor headers. The profile does not rely on direct access to the gated Meta base config." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5636526760540158, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-tgi-vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, Marlin/GEMM kernel differences, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by logical model parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Hugging Quants Llama 3.1 70B AWQ API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "serving", "commit_sha" ], "notes": "At commit 2123003760781134cfc31124aa6560a45b491fdf, the live API records a public non-gated Transformers text-generation repo with llama, llama-3.1, autoawq, text-generation-inference, endpoints_compatible, 4-bit, awq, deploy:azure, region:us, and Llama 3.1 license metadata. Current downloads are 282835. The API does not include a safetensors parameter summary for this repo, so tensor bytes are audited from the safetensors index and shard headers." }, { "label": "Hugging Quants Llama 3.1 70B AWQ model card", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "serving" ], "notes": "The card states this is a community quantized version of meta-llama/Meta-Llama-3.1-70B-Instruct, quantized with AutoAWQ from FP16 to INT4 using GEMM kernels, zero-point quantization, and group size 128. It says roughly 35 GiB of VRAM are needed to load the checkpoint before KV cache/CUDA graphs and gives Transformers, AutoAWQ, TGI, and vLLM usage examples." }, { "label": "Hugging Quants Llama 3.1 70B AWQ served config", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens" ], "notes": "The config records LlamaForCausalLM with torch_dtype float16, hidden_size 8192, intermediate_size 28672, 80 layers, 64 attention heads, 8 KV heads, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Meta Llama 3.1 70B Instruct base API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-70B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "logical_parameter_count", "license" ], "notes": "The current base API records gated manual access at commit 1605565b47bb9346c5515c34102e054115b4f98b, text-generation pipeline, Llama 3.1 license, region:us, base_model metadata for meta-llama/Llama-3.1-70B, and BF16 safetensors total 70553706496 parameters. Raw base config and tensor index remain inaccessible in this audit environment, so the served AWQ config is the architecture source of truth." }, { "label": "Hugging Quants Llama 3.1 70B AWQ safetensors index and shard headers", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 39767785472 bytes across nine shards, matching direct range-read shard header spans. Headers contain 1843 tensors totaling 39.767785472 GB: I32 34.492907520 GB and F16 5.274877952 GB. Stored suffix totals are qweight 34.225520640 GB, qzeros 0.267386880 GB, scales 1.069547520 GB, and F16 weight tensors 4.205330432 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 8192] and contribute 1.050673152B parameters / 2.101346304 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 37.666439168 GB." }, { "label": "Hugging Quants Llama 3.1 70B AWQ linked-object HEAD checks", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/tree/2123003760781134cfc31124aa6560a45b491fdf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for all nine safetensors shards found linked sizes totaling 39767996256 bytes. The linked file sizes include safetensors JSON header/container overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served AWQ config, gated base API metadata, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and local scrape row." }, "notes": "Use this profile for the Hugging Quants AWQ INT4 artifact only. It is audited from the public served artifact and intentionally does not turn the gated Meta BF16 base profile into an audited profile." }, { "id": "hugging-quants--meta-llama-3-1-8b-instruct-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4", "title": "Hugging Quants Meta Llama 3.1 8B Instruct AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the Hugging Quants AutoAWQ INT4 package of Meta Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense-awq", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Model card lineage, served AWQ config, existing Llama 3.1 quantized config comparison, and direct safetensors header review", "config_compatible": false, "notes": "The model card identifies this as a community quantized version of meta-llama/Meta-Llama-3.1-8B-Instruct, quantized from FP16 to INT4 with AutoAWQ. The base repo raw config is gated in this audit environment, so direct base compatibility is not asserted. This profile audits the served Hugging Quants AWQ artifact directly from its public config and tensor headers, with the existing audited RedHatAI Llama 3.1 8B Instruct FP8 profile used only as supporting geometry context." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 5.727854592, "swept_weight_gb": 4.67718144, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "logical Llama 3.1 8B parameters represented by safetensors qweight plus F16 tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors with F16 scales and unquantized F16 embedding/head/norm tensors. Logical parameter counts are reconstructed from the served tensor layout: I32 qweight tensors unpack 8x to logical weights, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 32 layers, 8 KV heads, hidden size 4096, 32 attention heads, Llama 3 RoPE scaling, 131072 max positions, and no sliding-window setting, so this profile charges full-context K and V streams for ordinary cached text decode." }, "notes": "Dense LlamaForCausalLM AWQ profile using the served Hugging Quants repo config and tensor headers. The profile does not rely on direct access to the gated Meta base config." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7132837170679307, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, GEMM kernel differences, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Hugging Quants Llama 3.1 8B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "downloads", "serving", "commit_sha", "pipeline", "license" ], "notes": "At commit db1f81ad4b8c7e39777509fac66c652eb0a52f91, the live API records a public non-gated text-generation repo with transformers, safetensors, llama, llama-3.1, autoawq, 4-bit, awq, deploy:azure, endpoints_compatible, and region:us tags. Current downloads are 194297. The API exposes LlamaForCausalLM/model_type llama plus AWQ bits 4 metadata, but not a safetensors parameter summary." }, { "label": "Hugging Quants Llama 3.1 8B Instruct AWQ model card", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/raw/db1f81ad4b8c7e39777509fac66c652eb0a52f91/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "runtime_format", "serving" ], "notes": "The card states this is a community-driven quantized version of meta-llama/Meta-Llama-3.1-8B-Instruct, quantized using AutoAWQ from FP16 down to INT4 using GEMM kernels, zero-point quantization, and group size 128. It documents Transformers, AutoAWQ, TGI, and vLLM serving paths, and notes about 4 GiB VRAM is needed to load the checkpoint before KV cache and CUDA graphs." }, { "label": "Hugging Quants Llama 3.1 8B Instruct AWQ served config", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/raw/db1f81ad4b8c7e39777509fac66c652eb0a52f91/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens", "embedding_layout" ], "notes": "The config records LlamaForCausalLM, torch_dtype float16, hidden_size 4096, intermediate_size 14336, 32 layers, 32 attention heads, 8 KV heads, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "RedHatAI Llama 3.1 8B Instruct FP8 config comparison", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "base_model_proof" ], "notes": "Manual comparison against the already audited RedHatAI quantized sibling found matching checked geometry fields: LlamaForCausalLM, hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, 131072 max positions, Llama 3 RoPE scaling, rope_theta 500000, vocab size 128256, tie_word_embeddings false, and no sliding-window setting. The two artifacts differ in quantization format and runtime dtype metadata." }, { "label": "Hugging Quants Llama 3.1 8B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/raw/db1f81ad4b8c7e39777509fac66c652eb0a52f91/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 5727854592 bytes across two shards, matching direct range-read shard header spans. Headers contain 739 tensors totaling 5.727854592 GB: I32 3.516923904 GB and F16 2.210930688 GB. Stored suffix totals are qweight 3.489660928 GB, qzeros 0.027262976 GB, scales 0.109051904 GB, and F16 weight tensors 2.101878784 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 4096] and contribute 0.525336576B parameters / 1.050673152 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 4.677181440 GB." }, { "label": "Hugging Quants Llama 3.1 8B Instruct AWQ linked-object HEAD checks", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/tree/db1f81ad4b8c7e39777509fac66c652eb0a52f91", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for both safetensors shards found linked sizes 4.677265296 GB and 1.050673280 GB. The linked file sizes include safetensors JSON header/container overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served AWQ config, existing audited Llama 3.1 quantized config comparison, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and local scrape row." }, "notes": "This profile supersedes the scraped 4 GB flat estimate by using exact stored AWQ bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "huihui-ai--huihui-deepseek-v4-flash-abliterated-ds4-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF", "title": "Huihui DeepSeek V4 Flash Abliterated GGUF Q2_K DS4", "summary": "Audited memory-side bounds profile for the API-selected Huihui Q2_K DS4 GGUF package of DeepSeek V4 Flash.", "model_family": "deepseek-v4-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V4-Flash", "relation": "quantized", "source": "Hugging Face model card/API base_model metadata, DeepSeek V4 Flash config, and selected GGUF header metadata", "config_compatible": true, "notes": "The Huihui repo identifies this package as a quantized/abliterated derivative of deepseek-ai/DeepSeek-V4-Flash. The selected GGUF header records the same DeepSeek V4 Flash geometry used by the base config." }, "architecture": { "canonical_architecture_id": "deepseek-v4-flash", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 99.705676384, "main_resident_weight_gb": 98.641279324, "auxiliary_resident_weight_gb": 1.06439706, "fixed_weight_gb": 7.742323036, "routed_expert_weight_gb": 0.355074048, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "API-selected Huihui Q2_K DS4 GGUF linked file size and range-read tensor index", "traffic_scope": "ordinary DS4 decode through Q2_K GGUF tensors, excluding the resident-only input embedding and GGUF metadata", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for each ordinary decode token", "shared_expert_notes": "The DeepSeek config and GGUF header record one shared expert. Shared expert tensors are stored outside blk.*.ffn_*_exps.weight and are charged in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "The HF API gguf.totalFileSize selects Huihui-DeepSeek-V4-Flash-BF16-abliterated-ds4-Q2_K.gguf. The model card recommends the smaller Q2 artifact for 128 GB Mac machines, but this profile follows the repo-level selected-artifact rule and records Q2_K. The optional MTP sidecar is not included." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00314, "notes": "V1 compressed-state coefficient from the original DeepSeek V4 Flash DS4 bounds worked example." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00121875, "notes": "V1 compressed-state read coefficient chosen to match the original DeepSeek V4 Flash DS4 worked-example ceiling." }, "notes": "DeepSeek V4 uses compressed/sparse long-context attention. Bounds Engine v1 uses the same audited DS4 compressed-state coefficients as the antirez DeepSeek V4 GGUF profile rather than a generic KV ratio." }, "notes": "This profile targets the selected Huihui Q2_K DS4 GGUF artifact, not the card-recommended Q2 artifact or the optional MTP sidecar." }, "serving": { "weight_format": "q2_mixed", "weight_bytes_per_param": 0.35066322486499185, "kv_store_format": "ds4_compressed", "kv_store_bytes_per_scalar": 1, "kv_read_format": "ds4_compressed", "kv_read_bytes_per_scalar": 1, "runtime_format": "ds4-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact Q2_K GGUF tensor spans for decode traffic and linked artifact file size for residency. DS4 kernel, dequantization, graph, scheduler, offload, and speculative decoding overheads are outside Bounds Engine v1.", "notes": "The selected artifact uses mixed Q2_K, Q8_0, F16, F32, and I32 tensor classes. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Huihui DeepSeek V4 Flash Abliterated GGUF HF API metadata", "url": "https://huggingface.co/api/models/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit f06f59bce3c36b3282b75c9fe2621c83c9399d10 records a public non-gated MIT GGUF repo with base_model deepseek-ai/DeepSeek-V4-Flash, relation quantized, region:us, 584606 downloads, GGUF architecture deepseek4, 1048576 context length, gguf.total 284334567511, and gguf.totalFileSize 99705676384. The totalFileSize matches the Q2_K linked object." }, { "label": "Huihui DeepSeek V4 Flash Abliterated GGUF model card", "url": "https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF/blob/f06f59bce3c36b3282b75c9fe2621c83c9399d10/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "weight_format", "runtime_format", "artifact_selection_notes" ], "notes": "The card records MIT licensing, base_model deepseek-ai/DeepSeek-V4-Flash, DS4/llama.cpp-specific GGUF packaging, and a file table for Q2, IQ2_XXS, Q2_K, Q4_K, and optional MTP files. It recommends Q2 for 128 GB Mac machines and Q4 for machines with at least 256 GB RAM; this profile records the API-selected Q2_K artifact separately from that card recommendation." }, { "label": "DeepSeek V4 Flash HF API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Flash", "source_type": "derived_calculation", "supports": [ "base_model_proof", "source_precision", "source_params" ], "notes": "The live base API response at commit 60d8d70770c6776ff598c94bb586a859a38244f1 records a public non-gated MIT text-generation repo with DeepSeek V4 Flash safetensors parameters totaling 158069433298 logical source tensors and mixed BF16/F32/FP8/I8 metadata." }, { "label": "DeepSeek V4 Flash config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/config.json", "source_type": "config", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention" ], "notes": "The config records DeepseekV4ForCausalLM, BF16 source dtype, 43 layers, hidden size 4096, 64 attention heads, one KV head, 512 head dimension, 256 routed experts, 6 experts per token, one shared expert, 1048576 max position embeddings, and compression-ratio metadata." }, { "label": "Huihui DeepSeek V4 Flash linked-object HEAD checks", "url": "https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF/tree/f06f59bce3c36b3282b75c9fe2621c83c9399d10", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "sidecar_exclusion" ], "notes": "HEAD checks found Q2 86.720111200 GB, Q2_K 99.705676384 GB, Q4_K 164.633502304 GB, and the optional MTP sidecar 3.807602400 GB. The API gguf.totalFileSize exactly matches the Q2_K file, so this profile selects Q2_K and excludes the optional MTP sidecar." }, { "label": "Huihui DeepSeek V4 Flash Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF/resolve/f06f59bce3c36b3282b75c9fe2621c83c9399d10/Huihui-DeepSeek-V4-Flash-BF16-abliterated-ds4-Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 16MB range-read of the selected GGUF v3 header found 58 metadata entries and 1328 tensors. The file is 99.705676384 GB, with data starting at byte 5333536. Tensor spans sum to 99.700341084 GB; metadata/tokenizer/header/file overhead accounts for 0.005335300 GB. token_embd.weight is 1.059061760 GB resident-only F16. Non-embedding fixed decode tensors sum to 7.742323036 GB. Routed expert Q2_K tensors sum to 90.898956288 GB across 43 layers and 256 uniform experts, or 0.355074048 GB per expert index. Type spans are Q2_K 90.898956288 GB, Q8_0 6.598885376 GB, F16 2.191345664 GB, I32 0.009308160 GB, and F32 0.001845596 GB." }, { "label": "DS4 engine README", "url": "https://github.com/antirez/ds4", "source_type": "vendor_doc", "supports": [ "runtime_format", "compressed_state" ], "notes": "The DS4 README identifies DS4-specific DeepSeek V4 Flash GGUF packages and runtime behavior. Bounds Engine v1 keeps DS4 runtime overhead outside the memory-side profile and uses audited compressed-state coefficients for the bounds calculation." }, { "label": "Original local frontier bounds note", "url": "https://github.com/osolmaz/onurclaw/blob/main/docs/2026-06-30-local-frontier-model-bounds.md", "source_type": "manual_review", "supports": [ "worked_example_parameters", "bounds_regression_target" ], "notes": "This evidence preserves the original rounded DeepSeek V4 DS4 worked-example target. The production profile uses exact selected Huihui Q2_K artifact bytes while reusing the same DS4 compressed-state allocation/read coefficients as the antirez profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned Huihui model card, pinned DeepSeek V4 Flash config, linked-object HEAD checks, a direct selected-Q2_K GGUF header/tensor-index range read, DS4 documentation, and the original bounds-note context." }, "notes": "Use this profile only for the Huihui API-selected Q2_K DS4 GGUF artifact. The card-recommended Q2 artifact, Q4_K artifact, and optional MTP sidecar have different resident and traffic bytes and require separate selected profiles if used." }, { "id": "hyper-ai--qwen3-5-9b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Hyper-AI/Qwen3.5-9B-fp8", "title": "Hyper-AI Qwen3.5 9B LightVL FP8", "summary": "Audited memory-side text-decode bounds profile for the Hyper-AI LightVL FP8 compressed-tensors package of Qwen3.5 9B Base.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B-Base", "relation": "quantized", "source": "Hugging Face API metadata, pinned LightVL compressed-tensors config, model card metadata, base profile comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The API and model card metadata identify Qwen/Qwen3.5-9B-Base as the quantized base model. Manual comparison found the same checked memory-relevant text geometry as the audited base profile: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, MTP tensors, untied embeddings, and 262144 max position embeddings. The package adds LightVL compressed-tensors FP8 metadata while preserving the base hybrid attention geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 12.147390432, "swept_weight_gb": 8.957726208, "auxiliary_resident_weight_gb": 3.189664224, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct LightVL compressed-tensors FP8/BF16/F32 safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Logical parameter counts follow the audited Qwen3.5 9B Base profile so model identity remains the 9.653104368B architecture. Byte fields use exact stored tensor payloads from the four safetensors shard headers. This LightVL package quantizes language self-attention, language MLP, language linear-attention matrix weights, and MTP matrix weights to F8_E4M3 while keeping embeddings, lm_head, vision tensors, layer norms, biases, selected scales, and recurrent-state runtime storage outside the FP8 matrix payload." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The LightVL compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, dequantization, dynamic activation quantization, and runtime scheduler behavior remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The config records kv_cache_scheme null, so KV cache is charged at BF16." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.2583921160397424, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-lightvl-compressed-tensors-fp8-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors safetensors bytes: F8_E4M3 matrix weights plus BF16/F32 side tensors and unquantized modules. Dynamic per-token activation quantization, dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors float quantization with static per-channel FP8 weights, dynamic per-token FP8 input activations, and kv_cache_scheme null. The ignore list excludes lm_head, language embeddings, and visual modules from FP8 quantization, while header grouping shows language linear-attention matrix weights and MTP matrix weights are also stored as F8_E4M3 in this package." }, "evidence": [ { "label": "Hyper-AI Qwen3.5 9B FP8 API metadata", "url": "https://huggingface.co/api/models/Hyper-AI/Qwen3.5-9B-fp8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving" ], "notes": "At repo SHA 83d256364f0daeacbcb0ce02c39d3a3101752477, the live API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B-Base, with fp8, quant, lightvl, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 104742. The API safetensors block reports BF16: 2492800496, F8_E4M3: 7161774080, F32: 3840, and total: 9654578416 storage-accounting elements." }, { "label": "Hyper-AI Qwen3.5 9B FP8 model card", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "quantization" ], "notes": "The card metadata records base_model Qwen/Qwen3.5-9B-Base and tags vlm, fp8, quant, lightvl, and qwen3.5. The card text states a 19G to 12G memory decrease, claims vLLM serving support, and describes LightVL as a VLM quantization toolkit supporting FP8, INT8, and FP8-Block." }, { "label": "Hyper-AI Qwen3.5 9B FP8 config", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8/raw/83d256364f0daeacbcb0ce02c39d3a3101752477/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, compressed-tensors float quantization, static per-channel FP8 weights, dynamic per-token FP8 activations, text_config model_type qwen3_5_text, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision_config, and kv_cache_scheme null." }, { "label": "Qwen3.5 9B Base audited profile comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state", "logical_parameter_counts" ], "notes": "Manual comparison against the audited Qwen3.5 9B Base profile found matching checked top-level and text geometry fields. The logical resident, swept, and auxiliary parameter counts in this profile follow the audited base architecture, while weight traffic follows this quantized artifact's exact stored bytes." }, { "label": "Hyper-AI Qwen3.5 9B FP8 safetensors headers", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8/raw/83d256364f0daeacbcb0ce02c39d3a3101752477/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 12147390432 bytes across four safetensors shards. Direct range reads of all shard headers found 1031 tensors totaling exactly 12.147390432 GB: BF16 4.985600992 GB, F8_E4M3 7.161774080 GB, and F32 0.000015360 GB. Linked-object HEAD checks resolved the four shards to 12.147517240 GB, leaving 0.000126808 GB of safetensors header/container overhead outside tensor payloads. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 8.957726208 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.189664224 GB. Header buckets are language MLP tensors 4.833673216 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, language linear-attention tensors 1.619283456 GB, visual 0.912020960 GB, language self-attention tensors 0.469999616 GB, MTP 0.243405824 GB, language layer/norm tensors 0.000524288 GB, and final language norm 0.000008192 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card metadata, pinned LightVL compressed-tensors config, direct safetensors shard-header range reads, linked-object HEAD checks, base Qwen3.5 9B profile comparison, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 1-byte dense weights and all-layer full-context KV. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept text-decode weights and includes the fixed DeltaNet runtime-state charge." }, { "id": "inclusionai--llada2-0-mini", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "inclusionAI/LLaDA2.0-mini", "title": "inclusionAI LLaDA2.0 Mini BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 LLaDA2.0 Mini block-diffusion MoE repo.", "model_family": "llada2-block-diffusion-moe", "base_model_proof": { "base_model": "inclusionAI/LLaDA2.0-mini", "relation": "base", "source": "Hugging Face API metadata, model card, pinned config, and custom Transformers modeling code", "config_compatible": true, "notes": "This profile represents the public base LLaDA2.0 Mini checkpoint itself. The repo uses custom LLaDA2MoeModelLM Transformers code rather than a standard autoregressive decoder implementation." }, "architecture": { "canonical_architecture_id": "llada2-0-mini", "max_context_tokens": 32768, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.511286784, "main_resident_weight_gb": 32.511286784, "auxiliary_resident_weight_gb": 0, "fixed_weight_gb": 1.146281472, "routed_expert_weight_gb": 0.119537664, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "Exact tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because LLaDA2 block-wise diffusion is not ordinary one-output-token autoregressive decode.", "shared_expert_notes": "The config records num_shared_experts 1, num_experts_per_tok 8, 256 routed experts, and first_k_dense_replace 1. The first layer is dense; layers 1-19 have one shared expert plus routed experts.", "notes": "Header-derived bytes are used. The seven safetensors shards contain 14,813 BF16 tensors totaling 32.511286784 GB. model.word_embeddings.weight is 0.643825664 GB and resident input embedding storage. lm_head.weight is 0.643825664 GB. model.norm.weight is 0.000004096 GB. Non-routed layer tensors, including attention, norms, dense first-layer MLP, and shared expert tensors, total 0.502451712 GB. Fixed non-embedding traffic is therefore 1.146281472 GB. Routed expert tensors total 30.601641984 GB and divide exactly into 256 uniform expert groups of 0.119537664 GB." }, "kv_adapter": { "kind": "unknown", "reason": "LLaDA2.0 Mini uses block-wise diffusion over masked output tokens with iterative denoising steps, not normal one-output-token autoregressive decode. Bounds Engine v1 only models ordinary autoregressive per-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "The pinned config sets use_cache false. The custom generate path builds a block-diffusion attention mask, initializes generated positions as mask tokens, then loops over blocks and denoising steps while repeatedly calling forward on the current block/window. A production profile needs a dedicated LLaDA2 diffusion adapter with block length, step count, mask schedule, block attention cost, cache policy, and MoE routing traffic." }, "notes": "This profile intentionally fails closed even though config and tensor headers are accessible, because the supported comparison math does not model LLaDA2's block-wise diffusion generation algorithm." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-llada2-block-diffusion-transformers-bf16", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The Hugging Face API safetensors metadata and direct safetensors headers record only BF16 tensor payloads." }, "evidence": [ { "label": "inclusionAI LLaDA2.0 Mini Hugging Face API metadata", "url": "https://huggingface.co/api/models/inclusionAI/LLaDA2.0-mini", "source_type": "derived_calculation", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit dad945cac317da394b390f82c7b40691d8a881ed, the API reports a public non-gated Apache-2.0 Transformers text-generation repo with llada2_moe, dllm, diffusion, custom_code, region:us, 146,675 downloads, and BF16 safetensors parameters totaling 16.255643392B." }, { "label": "inclusionAI LLaDA2.0 Mini served config", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "unsupported_reason" ], "notes": "The config records LLaDA2MoeModelLM, model_type llada2_moe, hidden size 2048, 20 layers, 16 attention heads, 4 KV heads, 128 head dimension, 32768 max positions, sliding_window 4096 with use_sliding_window false, use_cache false, vocab size 157184, 256 experts, 8 routed experts per token, one shared expert, and first_k_dense_replace 1." }, { "label": "inclusionAI LLaDA2.0 Mini model card", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm", "active_params_b" ], "notes": "The model card describes LLaDA2.0 Mini as a diffusion language model with 16B total non-embedding parameters and about 1.4B activated during inference. Its sample generation call uses gen_length, block_length 32, and steps 32, which is not a standard autoregressive decode loop." }, { "label": "inclusionAI LLaDA2.0 Mini custom modeling code", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/modeling_llada2_moe.py", "source_type": "manual_review", "supports": [ "unsupported_reason", "generation_algorithm", "kv_adapter" ], "notes": "The custom generate method says it operates differently from standard autoregressive generation. It creates a full masked output template, constructs a block-diffusion attention mask, and repeatedly denoises mask tokens block by block across a configurable number of steps." }, { "label": "inclusionAI LLaDA2.0 Mini safetensors index and shard headers", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 32511286784 bytes across seven shards. Direct range-read safetensors headers found 14,813 BF16 tensors totaling the same 32.511286784 GB. Linked shard files total 32.513122504 GB, leaving 0.001835720 GB of safetensors header/container bytes outside tensor payloads. model.word_embeddings.weight is 0.643825664 GB, lm_head.weight is 0.643825664 GB, model.norm.weight is 0.000004096 GB, non-routed layer tensors total 0.502451712 GB, and routed expert tensors total 30.601641984 GB. The 256 expert groups are uniform at 0.119537664 GB each." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from current HF API metadata, pinned model card, pinned config, custom Transformers modeling code, safetensors index, expanded linked-object sizes, and direct shard header byte grouping. Marked unsupported because Bounds Engine v1 lacks a LLaDA2 block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model LLaDA2 block-wise diffusion over masked output tokens, so ordinary autoregressive throughput would be misleading even though resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated LLaDA2 block-diffusion adapter exists." }, { "id": "ista-daslab--gemma-3-27b-it-gptq-4b-128g", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g", "title": "ISTA-DASLab Gemma 3 27B IT GPTQ 4B 128G", "summary": "Audited memory-side text-decode bounds profile for the ISTA-DASLab GPTQ INT4 Gemma 3 27B IT serving artifact.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-27b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata and the served ISTA-DASLab config", "config_compatible": false, "notes": "The model card/API metadata identify google/gemma-3-27b-it as the base model, but that base repo remains gated in this audit environment. This profile therefore uses the public ISTA-DASLab served config and safetensors headers directly instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.632394708, "swept_params_b": 27.209334372, "auxiliary_resident_params_b": 0.423060336, "resident_weight_gb": 16.867252224, "swept_weight_gb": 16.021131552, "auxiliary_resident_weight_gb": 0.846120672, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "ordinary text decode charges quantized language layer tensors, language norm, and the tied BF16 language_model.model.embed_tokens.weight output projection", "auxiliary_scope": "vision_tower and multi_modal_projector tensors are resident for the multimodal package but not swept for ordinary text decode", "notes": "Range-read safetensors headers record 2115 tensors totaling 27.632394708B logical parameters and 16.867252224 GB payload bytes. The checkpoint stores packed I32 int4 tensors, BF16 unquantized tensors, and tiny I64 side tensors. It does not store a separate language_model.lm_head.weight tensor, so language_model.model.embed_tokens.weight is charged in swept ordinary decode as the tied output projection. Resident-only non-text tensors are the BF16 vision tower plus multimodal projector." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window_pattern 6 over 62 language layers. Using the documented Gemma 3 pattern of five local layers followed by one global layer gives 10 full-context global layers." }, { "kind": "sliding_window", "layers": 52, "kv_heads": 16, "head_dim": 128, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 52 language layers use the config's 1024-token local sliding-window attention." } ], "notes": "Layered KV models ordinary text decode after any image prefill. The config records cache_implementation hybrid and kv_cache_scheme null, so the profile charges BF16 K and V streams." }, "notes": "Gemma3ForConditionalGeneration is multimodal. This profile models ordinary text decode, not vision encoder or image-prefill throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6104158688467444, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-gptq-int4-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 ignored modules and scale tensors, and I64 side tensors from safetensors headers. GPTQ dequantization, activation traffic, compute throughput, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The model card describes INT4 weight quantization with a symmetric per-group scheme, group size 128, and GPTQ. The served config records compressed-tensors pack-quantized int4 weights, bfloat16 model dtype, and kv_cache_scheme null; this profile therefore charges exact stored tensor bytes for weights and BF16 KV cache traffic." }, "evidence": [ { "label": "ISTA-DASLab Gemma 3 27B GPTQ API metadata", "url": "https://huggingface.co/api/models/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b, the live API reports a public non-gated Gemma-licensed image-text-to-text repo with base_model google/gemma-3-27b-it, INT4/vLLM/llmcompressor/compressed-tensors tags, region:us, 236301 downloads, and safetensors logical parameters I64: 868, I32: 25598361600, BF16: 2034032240, total: 27632394708. The model card identifies the artifact as a quantized derivative of google/gemma-3-27b-it, with INT4 grouped GPTQ quantization." }, { "label": "ISTA-DASLab Gemma 3 27B GPTQ model card", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving" ], "notes": "The card says the model was obtained by quantizing google/gemma-3-27b-it weights to INT4, says only linear operators within language_model transformer blocks are quantized, and says the vision model plus multimodal projection are kept in original precision. It also states symmetric per-group quantization with group size 128 and GPTQ." }, { "label": "ISTA-DASLab Gemma 3 27B GPTQ config", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/raw/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "sliding_window_pattern", "max_context_tokens", "serving", "vision_geometry" ], "notes": "The config records Gemma3ForConditionalGeneration, gemma3_text, bfloat16 model dtype, compressed-tensors pack-quantized int4 weights, group_size 128, kv_cache_scheme null, cache_implementation hybrid, 62 text layers, 5376 hidden size, 21504 intermediate size, 32 attention heads, 16 KV heads, 128 head dimension, 131072 max positions, 1024-token sliding window, sliding_window_pattern 6, and a 27-layer SigLIP vision tower. The quantization ignore list excludes language_model.lm_head, vision_tower, and multi_modal_projector, but the safetensors headers contain no separate language_model.lm_head tensor." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes the repeating local/global attention pattern as five local attention layers followed by one global attention layer, and says larger Gemma 3 models support 128k context. Applied to the 62 layers in the served config, that yields 10 global layers and 52 local layers." }, { "label": "ISTA-DASLab Gemma 3 27B GPTQ safetensors index", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/raw/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "total_params_b", "weight_format" ], "notes": "The safetensors index maps tensors across five shards but does not include total_size metadata. Direct header reads found tensor payload bytes totaling 16.867252224 GB." }, { "label": "ISTA-DASLab Gemma 3 27B GPTQ safetensors headers", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/tree/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_scope", "weight_format" ], "notes": "Range-reads of all five safetensors shard headers found 2115 tensors totaling 16.867252224 GB: I32 packed tensors total 12.799180800 GB, BF16 tensors total 4.068064480 GB, and I64 side tensors total 0.000006944 GB. Logical parameters total 27.632394708B after expanding weight_packed I32 tensors by 8. Ordinary swept text traffic totals 27.209334372B logical parameters / 16.021131552 GB: input/output-tied language_model.model.embed_tokens.weight 2.819260416 GB, language MLP 11.087293344 GB, language self-attention 2.111900544 GB, other language layer tensors 0.002666496 GB, and language norm 0.000010752 GB. Resident-only vision_tower tensors total 0.833732064 GB and multi_modal_projector tensors total 0.012388608 GB. Linked-object HEAD checks resolved the five shards to 16.867530008 GB, leaving 277784 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Google Gemma 3 27B IT gated base access note", "url": "https://huggingface.co/google/gemma-3-27b-it", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base repo remains gated in this audit environment, matching the existing google/gemma-3-27b-it unsupported profile. This audited ISTA-DASLab profile does not copy base geometry; it relies on the public derivative config and tensor headers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card metadata, served config, generation config, safetensors index, direct safetensors shard header range reads, and Gemma 3 local/global attention documentation." }, "notes": "This profile supersedes the generated row's rounded ideal 4-bit estimate with exact compressed-tensors GPTQ INT4 stored bytes and a Gemma 3 layered local/global KV adapter." }, { "id": "jackrong--qwen3-5-27b-claude-4-6-opus-reasoning-distilled-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF", "title": "Jackrong Qwen3.5 27B Claude Opus Reasoning Distilled GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the Q4_K_M GGUF artifact of Jackrong's Qwen3.5 27B Claude Opus reasoning-distilled model.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo metadata records Qwen/Qwen3.5-27B as the base model and quantized base. Manual comparison against the pinned Qwen base config found matching memory-relevant top-level, text, layer-pattern, and vision geometry fields. The selected GGUF header records the same Qwen3.5 text geometry and excludes the separate multimodal projector sidecar." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 16.540267968, "swept_weight_gb": 15.814117376, "auxiliary_resident_weight_gb": 0.726150592, "resident_parameter_scope": "selected Qwen3.5-27B.Q4_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj-BF16.gguf sidecar is not included unless explicitly loaded", "notes": "The profile targets Qwen3.5-27B.Q4_K_M.gguf because the model card's hardware/speed note is for Q4_K_M. The live HF API gguf.totalFileSize currently matches Qwen3.5-27B.Q2_K.gguf instead. The selected linked file is 16.540267968 GB. Header tensor spans total 16.529278976 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010988992 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, and ordinary blk.0-63 tensors, with no MTP, mmproj, vision, or draft tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and selected GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate mmproj-BF16.gguf sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6149713307776606, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-qwen3.5-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected profile artifact is Qwen3.5-27B.Q4_K_M.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwen3.5 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "The live HF API response at commit 5825aa2c6c361e3ee4f588021192d1f919dd60d7 records a public Apache-2.0 GGUF repo with base_model Qwen/Qwen3.5-27B, 212029 downloads, region:us, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 10121835936. The API totalFileSize matches Qwen3.5-27B.Q2_K.gguf, while this profile targets Q4_K_M because the card's reported hardware/speed note is for Q4_K_M." }, { "label": "Jackrong Qwen3.5 27B GGUF model card", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/raw/5825aa2c6c361e3ee4f588021192d1f919dd60d7/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "serving" ], "notes": "The pinned card records Apache-2.0 licensing, base_model Qwen/Qwen3.5-27B, Unsloth fine-tuning, image-text-to-text metadata, and a community-tested hardware note stating about 16.5 GB VRAM with Q4_K_M quantization and 29-35 tok/s generation speed." }, { "label": "Jackrong Qwen3.5 27B GGUF config", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/raw/5825aa2c6c361e3ee4f588021192d1f919dd60d7/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The served config records Qwen3_5ForConditionalGeneration, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and false tied embeddings." }, { "label": "Qwen3.5 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no memory-relevant architecture differences between the served Jackrong config and the pinned base Qwen3.5 27B config. The target config makes dtype and false text tie_word_embeddings explicit where the base relies on defaults." }, { "label": "Jackrong Qwen3.5 27B GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/tree/5825aa2c6c361e3ee4f588021192d1f919dd60d7", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwen3.5-27B.Q2_K.gguf 10.121835936 GB, Qwen3.5-27B.Q3_K_M.gguf 13.289641376 GB, Qwen3.5-27B.Q3_K_S.gguf 12.073948576 GB, Qwen3.5-27B.Q4_K_M.gguf 16.540267968 GB, Qwen3.5-27B.Q4_K_S.gguf 15.568614848 GB, Qwen3.5-27B.Q8_0.gguf 28.595758528 GB, and mmproj-BF16.gguf 0.931145856 GB. The mmproj sidecar is not part of the selected main text artifact." }, { "label": "Jackrong Qwen3.5 27B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/resolve/5825aa2c6c361e3ee4f588021192d1f919dd60d7/Qwen3.5-27B.Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 851 tensors. The linked file is 16.540267968 GB. Tensor spans sum to 16.529278976 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.715161600 GB, and ordinary blk.0-63 tensors 14.771152896 GB. Metadata/tokenizer/header/file overhead accounts for 0.010988992 GB. Tensor spans split into Q4_K 11.362959360 GB, Q5_K 1.730150400 GB, Q6_K 3.425587200 GB, and F32 0.010582016 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, and no MTP, mmproj, vision, or draft tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, served config, pinned Qwen base config comparison, linked GGUF file sizes, direct selected-GGUF header/tensor-index range read, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the Jackrong Q4_K_M main GGUF text artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or MTP speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "jackrong--qwen3-5-9b-deepseek-v4-flash-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF", "title": "Jackrong Qwen3.5 9B DeepSeek V4 Flash GGUF Q3_K_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q3_K_M GGUF artifact of Jackrong's Qwen3.5 9B DeepSeek V4 Flash distill.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "unsloth/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Unsloth base config, upstream Qwen config comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo records unsloth/Qwen3.5-9B as its base model. The Unsloth base repo records Qwen/Qwen3.5-9B as its base and keeps the Qwen3.5 9B text geometry. The selected GGUF header records the same Qwen3.5 9B block, attention, context, and DeltaNet state geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.953803264, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.01711872, "resident_weight_gb": 4.623521024, "swept_weight_gb": 4.1755136, "auxiliary_resident_weight_gb": 0.448007424, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected Q3_K_M text GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode; mmproj.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf because HF API gguf.totalFileSize exactly matches that artifact. The selected linked file is 4.623521024 GB. Header tensor spans total 4.612556800 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010964224 GB. The main GGUF contains output.weight, token_embd.weight, blk.* tensors, and output_norm.weight; it has no mmproj, vision, audio, MTP, or rope_freqs tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 layers with every fourth layer using full attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q3_K_M GGUF artifact after any multimodal prefill. It does not include the separate vision projector sidecar." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.516375096445273, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q3-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and MTP speculative decoding are outside Bounds Engine v1.", "notes": "The selected normal text artifact uses a mixed Q3_K_M layout: tensor payloads split into 2.299658240 GB Q3_K, 1.358954496 GB Q4_K, 0.115343360 GB Q5_K, 0.834355200 GB Q6_K, and 0.004245504 GB F32. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwen3.5 9B DeepSeek V4 Flash GGUF HF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit e842f7426b567ec1e335d5cc348f4c86cdcf1453 records a public Apache-2.0 image-text-to-text GGUF repo with base_model unsloth/Qwen3.5-9B, DeepSeek-V4 distillation tags, region:us, 181436 downloads, GGUF architecture qwen35, 262144 context length, gguf.total 8953803264, and gguf.totalFileSize 4623521024. The API totalFileSize matches Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf." }, { "label": "Jackrong Qwen3.5 9B DeepSeek V4 Flash GGUF model card", "url": "https://huggingface.co/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF/raw/e842f7426b567ec1e335d5cc348f4c86cdcf1453/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The card describes an efficient reasoning model distilled from DeepSeek-V4 data into the Qwen3.5 9B parameter space, records Apache-2.0 licensing, and references a Q5_K_M evaluation comparison. It does not declare a serving-default GGUF quantization, so this profile uses the artifact selected by HF API gguf.totalFileSize." }, { "label": "Unsloth Qwen3.5 9B base config", "url": "https://huggingface.co/unsloth/Qwen3.5-9B/raw/005429cee5cb648998cf2b70eebdd83175989c9a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and a resident vision config." }, { "label": "Qwen3.5 9B upstream config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison against Qwen/Qwen3.5-9B found the same Qwen3.5 9B text geometry and context fields as the Unsloth base config." }, { "label": "Jackrong Qwen3.5 9B DeepSeek V4 Flash GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF/tree/e842f7426b567ec1e335d5cc348f4c86cdcf1453", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found normal text files Q3_K_M 4.623521024 GB, Q4_K_M 5.629105408 GB, Q5_K_M 6.467966208 GB, Q6_K 7.359255808 GB, and Q8_0 9.527497984 GB. The separate mmproj.gguf sidecar is 0.921705376 GB. The selected profile uses Q3_K_M because it exactly matches API gguf.totalFileSize." }, { "label": "Jackrong Qwen3.5 9B DeepSeek V4 Flash Q3_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF/resolve/e842f7426b567ec1e335d5cc348f4c86cdcf1453/Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 40 metadata entries and 427 tensors. The linked file is 4.623521024 GB. Tensor spans sum to 4.612556800 GB: output.weight 0.834355200 GB, token_embd.weight 0.437043200 GB, blk.* tensors 3.341142016 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.010964224 GB. Tensor spans split into Q3_K 2.299658240 GB, Q4_K 1.358954496 GB, Q5_K 0.115343360 GB, Q6_K 0.834355200 GB, and F32 0.004245504 GB. The header records qwen35.block_count 32, context_length 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no mmproj/vision/audio/MTP tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, Unsloth base config, upstream Qwen config comparison, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q3_K_M artifact." }, "notes": "Use this profile for the API-selected Q3_K_M main text GGUF artifact. Do not infer multimodal projector residency unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "jackrong--qwopus3-5-9b-coder-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF", "title": "Jackrong Qwopus3.5 9B Coder MTP GGUF Q5_K_M", "summary": "Audited memory-side ordinary text-decode bounds profile for the selected Q5_K_M GGUF artifact of Jackrong's Qwopus3.5 9B Coder MTP package.", "model_family": "qwen3.5-dense-multimodal-coder-mtp", "base_model_proof": { "base_model": "Jackrong/Qwopus3.5-9B-v3.5", "relation": "derived_package", "source": "Hugging Face model card base_model metadata, Jackrong base config, Qwen3.5 base config comparison, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card identifies this Coder MTP package as a specialized model built on Jackrong/Qwopus3.5-9B-v3.5, itself an unsloth/Qwen3.5-9B derivative. The Jackrong served config matches the audited Qwen/Qwen3.5-9B memory-relevant geometry: Qwen3_5ForConditionalGeneration, 32 text layers, every fourth layer full attention, 4 KV heads, 256 full-attention head dimension, 24 linear-attention layers, resident vision config, and 262144 context. The selected GGUF header records the same qwen35 text geometry plus one MTP block." }, "architecture": { "canonical_architecture_id": "qwopus3-5-9b-coder-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.197093888, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.260409344, "resident_weight_gb": 6.642546592, "swept_weight_gb": 5.757732864, "auxiliary_resident_weight_gb": 0.884813728, "resident_parameter_scope": "selected Qwopus3.5-9B-Coder-MTP-Q5_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.31 tensors from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.32 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; mmproj-F32.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The HF API gguf.totalFileSize points at mmproj-F32.gguf with architecture clip, so API GGUF metadata is not a valid language-model artifact selector for this repo. The card does not name a main quantization for the benchmark table; this profile audits Q5_K_M as the concrete practical main artifact following the existing Jackrong MTP GGUF selected-artifact convention. The selected linked file is 6.642546592 GB. Header tensor spans total 6.631575552 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010971040 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, ordinary blk.0-31 tensors, and blk.32 MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.32 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Jackrong base config, Qwen3.5 base config, and selected GGUF metadata record 32 ordinary text layers with every fourth layer using full attention, giving eight full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the selected Q5_K_M GGUF artifact. The MTP draft block is resident for the package but does not reduce token traffic in Bounds Engine v1; multimodal projector use requires a separate workload profile." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7222440776283615, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-coder-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected profile artifact is Q5_K_M. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwopus3.5 Coder MTP GGUF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "api_artifact_caveat", "max_context_tokens" ], "notes": "The live HF API response at commit bf3fdbf18a25cf1ce449589b23c067ca29d6d1a9 records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model Jackrong/Qwopus3.5-9B-v3.5, 139234 downloads, region:us, text-generation-inference, qwen3_5, MTP, speculative-decoding, lora, sft, agent, coder, tool-use, and function-calling tags. The API gguf block reports architecture clip, gguf.total 456010480, and gguf.totalFileSize 921704480, which matches mmproj-F32.gguf rather than a language-model GGUF. This profile therefore does not use the API GGUF block as the text artifact selector." }, { "label": "Jackrong Qwopus3.5 Coder MTP GGUF model card", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/raw/bf3fdbf18a25cf1ce449589b23c067ca29d6d1a9/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope" ], "notes": "The pinned card records Apache-2.0 licensing, image-text-to-text packaging, base_model Jackrong/Qwopus3.5-9B-v3.5, agentic coding and tool-use specialization, and an MTP variant with draft=2. It reports GB10 benchmark data with concurrency 5 and overall throughput rising from 4.94 T/s to 6.71 T/s under MTP. The card points vision users to a separate mmproj GGUF file." }, { "label": "Jackrong Qwopus3.5 9B v3.5 config", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-v3.5/raw/dc2b00e1b1bc404133e3a3e15e7ddcdff814fd86/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, BF16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and resident vision config." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state", "kv_adapter" ], "notes": "Manual comparison found matching checked memory-relevant fields between Jackrong/Qwopus3.5-9B-v3.5 and Qwen/Qwen3.5-9B: architecture, model_type, tie_word_embeddings, text model_type, 32 text layers, full_attention_interval 4, hidden size, intermediate size, attention geometry, linear-attention geometry, mamba_ssm_dtype, vocabulary size, and 262144 max positions." }, { "label": "Jackrong Qwopus3.5 Coder MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/tree/bf3fdbf18a25cf1ce449589b23c067ca29d6d1a9", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwopus3.5-9B-Coder-MTP-Q5_K_M.gguf 6.642546592 GB, BF16 18.407317216 GB, IQ4_XS 5.357877152 GB, Q2_K 3.914971040 GB, Q3_K_L 5.048514464 GB, Q3_K_M 4.737611680 GB, Q3_K_S 4.364023712 GB, Q4_K_M 5.780092832 GB, Q4_K_S 5.488555936 GB, Q5_K_S 6.472644512 GB, Q6_K 7.558903712 GB, Q8_0 9.786062752 GB, and mmproj-F32.gguf 0.921704480 GB. The API gguf.totalFileSize matches mmproj-F32, while this profile audits Q5_K_M as the concrete selected main text artifact." }, { "label": "Jackrong Qwopus3.5 Coder MTP Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/bf3fdbf18a25cf1ce449589b23c067ca29d6d1a9/Qwopus3.5-9B-Coder-MTP-Q5_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 53 metadata entries and 442 tensors. The linked file is 6.642546592 GB. Tensor spans sum to 6.631575552 GB: output.weight plus output_norm.weight 0.834371584 GB, token_embd.weight 0.699269120 GB, ordinary blk.0-31 tensors 4.923361280 GB, and blk.32 MTP tensors 0.174573568 GB. Metadata/tokenizer/header/file overhead accounts for 0.010971040 GB. Tensor spans split into Q5_K 4.743495680 GB, Q6_K 1.883750400 GB, and F32 0.004329472 GB. The header records qwen35.block_count 33, context_length 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 4096, ssm.conv_kernel 4, nextn_predict_layers 1, and no mmproj or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned Jackrong base config, Qwen3.5 base config comparison, linked-object HEAD checks, direct GGUF header/tensor-index range read of the selected Q5_K_M artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the selected Jackrong Qwopus3.5 9B Coder MTP Q5_K_M GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer API-selected mmproj residency, other quantization bytes, multimodal projector execution, or MTP speculative acceleration unless a workload profile explicitly selects those paths." }, { "id": "jackrong--qwopus3-6-27b-coder-compat-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF", "title": "Jackrong Qwopus3.6 27B Coder Compat MTP GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the Compat card-selected Q4_K_M GGUF artifact of Jackrong's Qwopus3.6 27B Coder MTP package.", "model_family": "qwen3.6-dense-multimodal-coder-compat-mtp", "base_model_proof": { "base_model": "Jackrong/Qwopus3.6-27B-Coder", "relation": "derived_package", "source": "Hugging Face API/card base_model metadata, Jackrong Coder config, linked-object HEAD checks, selected GGUF header metadata, and existing Qwen3.6/Qwen3.5 runtime reviews", "config_compatible": true, "notes": "The Compat repo card identifies this package as a compatibility-template update of Jackrong/Qwopus3.6-27B-Coder with unchanged weights and bundled MTP tensors. The Jackrong Coder served config records the same Qwen3.6 memory-relevant geometry used by the existing Coder MTP profile: Qwen3_5ForConditionalGeneration, 64 text layers, every fourth layer full attention, 4 KV heads, 256 full-attention head dimension, 48 linear-attention layers, resident vision config, and 262144 context. The selected Q4_K_M GGUF header records the same qwen35 text geometry plus one MTP block." }, "architecture": { "canonical_architecture_id": "qwopus3-6-27b-coder-compat-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.320697856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.696097792, "resident_weight_gb": 16.810710112, "swept_weight_gb": 15.821244416, "auxiliary_resident_weight_gb": 0.989465696, "resident_parameter_scope": "selected Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.64 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; mmproj-F32.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf because the Compat-specific card section validates a weight-identical Q4_K_M template comparison across Hugging Face Jinja and llama.cpp/minja, reports Q4_K_M median decode speed, and records Q4_K_M MTP acceptance. The retained original Coder card section still reports a Q5_K_M SWE-bench run, but that inherited benchmark does not override the Compat update's Q4_K_M validation artifact. The HF API gguf.totalFileSize points at mmproj-F32.gguf with architecture clip, so API GGUF metadata is not a valid language-model artifact selector for this repo. The selected linked file is 16.810710112 GB. Header tensor spans total 16.799719424 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010990688 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, ordinary blk.0-63 tensors, and blk.64 MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.64 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Jackrong Coder config and selected GGUF metadata record 64 ordinary text layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the selected Q4_K_M GGUF artifact. The MTP draft block is resident for the package but does not reduce token traffic in Bounds Engine v1; multimodal projector use requires a separate workload profile." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6153104214469447, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-coder-compat-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected profile artifact is Q4_K_M because the Compat-specific card validation uses Q4_K_M. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwopus3.6 Coder Compat MTP GGUF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "api_artifact_caveat", "max_context_tokens" ], "notes": "The live HF API response at commit 2b74b86b39dda302b0cb165145d2b0a48c31e833 records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model Jackrong/Qwopus3.6-27B-Coder, 322097 downloads, region:us, text-generation-inference, qwen3_6, agent, tool-use, coder, mtp, speculative-decoding, compatibility, and multimodal tags. The API gguf block reports architecture clip, gguf.total 460730096, and gguf.totalFileSize 931147296, which matches mmproj-F32.gguf rather than a language-model GGUF. This profile therefore does not use the API GGUF block as the text artifact selector." }, { "label": "Jackrong Qwopus3.6 Coder Compat MTP GGUF model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/raw/2b74b86b39dda302b0cb165145d2b0a48c31e833/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope" ], "notes": "The pinned card records Apache-2.0 licensing, image-text-to-text GGUF packaging, base_model Jackrong/Qwopus3.6-27B-Coder, and a Compatibility Edition template update that accepts JSON-string, mapping, and list tool-call arguments across Hugging Face Jinja and llama.cpp/minja. The Compat-specific validation used two weight-identical Q4_K_M GGUF files across all 866 tensors, records HF rendering 10/10, minja rendering 10/10, median decode speed 30.53 tok/s, and MTP acceptance 76.81%. The retained original Coder card below records the earlier Q5_K_M SWE-bench run; this profile selects Q4_K_M from the Compat-specific validation section." }, { "label": "Jackrong Qwopus3.6 27B Coder config", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder/raw/1eee80aef82e96900a3d3aa88c302e0dbf0e5e49/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, mtp_num_hidden_layers 1, unsloth_fixed_mtp true, and resident vision config." }, { "label": "Jackrong Qwopus3.6 Coder Compat MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/tree/2b74b86b39dda302b0cb165145d2b0a48c31e833", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf 16.810710112 GB, BF16 54.657729632 GB, IQ4_XS 15.420445792 GB, Q2_K 10.864587872 GB, Q3_K_L 14.559794272 GB, Q3_K_M 13.500732512 GB, Q3_K_S 12.256531552 GB, Q4_K_S 15.825294432 GB, Q5_K_M 19.535696992 GB, Q5_K_S 18.971677792 GB, Q6_K 22.430995552 GB, Q8_0 29.047080032 GB, and mmproj-F32.gguf 0.931147296 GB. The API gguf.totalFileSize matches mmproj-F32, while the Compat-specific card validation selects Q4_K_M." }, { "label": "Jackrong Qwopus3.6 Coder Compat MTP Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/resolve/2b74b86b39dda302b0cb165145d2b0a48c31e833/Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 39 metadata entries and 866 tensors. The linked file is 16.810710112 GB. Tensor spans sum to 16.799719424 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.715161600 GB, ordinary blk.0-63 tensors 14.778279936 GB, and blk.64 MTP tensors 0.263313408 GB. Metadata/tokenizer/header/file overhead accounts for 0.010990688 GB. Tensor spans split into Q4_K 12.262440960 GB, Q6_K 4.526592000 GB, and F32 0.010686464 GB. The header records general.architecture qwen35, general.name Qwopus3.6 27B Coder Templatefix Hf, qwen35.block_count 65, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, nextn_predict_layers 1, and no mmproj or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned Jackrong Coder config, linked-object HEAD checks, direct GGUF header/tensor-index range read of the Compat-selected Q4_K_M artifact, and the existing Qwen3.5/Qwen3.6 runtime review." }, "notes": "Use this profile for the Jackrong Coder Compat MTP Q4_K_M GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer API-selected mmproj residency, Q5_K_M inherited benchmark bytes, other quantization bytes, multimodal projector execution, or MTP speculative acceleration unless a workload profile explicitly selects those paths." }, { "id": "jackrong--qwopus3-6-27b-coder-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF", "title": "Jackrong Qwopus3.6 27B Coder MTP GGUF Q5_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-benchmark Q5_K_M GGUF artifact of Jackrong's Qwopus3.6 27B Coder MTP package.", "model_family": "qwen3.6-dense-multimodal-coder-mtp", "base_model_proof": { "base_model": "Jackrong/Qwopus3.6-27B-v2", "relation": "derived_package", "source": "Hugging Face model card base_model metadata, Jackrong base config, Qwen3.6 base config comparison, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card identifies this Coder MTP package as a specialized model built on Jackrong/Qwopus3.6-27B-v2, itself a Qwen3.6-27B derivative. The Jackrong v2 served config matches the audited Qwen/Qwen3.6-27B memory-relevant geometry: Qwen3_5ForConditionalGeneration, 64 text layers, every fourth layer full attention, 4 KV heads, 256 full-attention head dimension, 48 linear-attention layers, resident vision config, and 262144 context. The selected GGUF header records the same qwen35 text geometry plus one MTP block." }, "architecture": { "canonical_architecture_id": "qwopus3-6-27b-coder-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.320697856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.696097792, "resident_weight_gb": 19.535698496, "swept_weight_gb": 18.346018816, "auxiliary_resident_weight_gb": 1.18967968, "resident_parameter_scope": "selected Qwopus3.6-27B-Coder-MTP-Q5_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.64 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; mmproj-F32.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets Qwopus3.6-27B-Coder-MTP-Q5_K_M.gguf because the model card's completed SWE-bench and speed report explicitly use Q5_K_M on RTX 5090 with MTP enabled. The HF API gguf.totalFileSize points at mmproj-F32.gguf with architecture clip, so API GGUF metadata is not a valid language-model artifact selector for this repo. The selected linked file is 19.535698496 GB. Header tensor spans total 19.524706304 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010992192 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, ordinary blk.0-63 tensors, and blk.64 MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.64 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Jackrong base config, Qwen3.6 base config, and selected GGUF metadata record 64 ordinary text layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the selected Q5_K_M GGUF artifact. The MTP draft block is resident for the package but does not reduce token traffic in Bounds Engine v1; multimodal projector use requires a separate workload profile." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.715051225959431, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-coder-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected profile artifact is Q5_K_M because the model card's published coding benchmark and speed claim use Q5_K_M. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwopus3.6 Coder MTP GGUF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "api_artifact_caveat", "max_context_tokens" ], "notes": "The live HF API response at commit ab9175af03e34068fd773eeb64c80461bec2a526 records a public non-gated Apache-2.0 text-generation GGUF repo with base_model Jackrong/Qwopus3.6-27B-v2, 297857 downloads, region:us, text-generation-inference, qwen3_6, lora, sft, agent, tool-use, and coder tags. The API gguf block reports architecture clip, gguf.total 460730096, and gguf.totalFileSize 931147296, which matches mmproj-F32.gguf rather than a language-model GGUF. This profile therefore does not use the API GGUF block as the text artifact selector." }, { "label": "Jackrong Qwopus3.6 Coder MTP GGUF model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope" ], "notes": "The pinned card records Apache-2.0 licensing, text-generation packaging, base_model Jackrong/Qwopus3.6-27B-v2, agentic coding and tool-use specialization, a Coder MTP variant, and a completed SWE-bench Verified full-500 run using Q5_K_M GGUF on RTX 5090 with MTP enabled at about 100 tokens/sec. This explicit benchmark setup selects Q5_K_M for the audited main text artifact." }, { "label": "Jackrong Qwopus3.6 27B v2 config", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2/raw/d0d82f4ccc9d41d4fe9595e96be4595327bb5de7/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and resident vision config. Checked fields match the audited Qwen/Qwen3.6-27B base config." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state", "kv_adapter" ], "notes": "Manual comparison found matching checked memory-relevant fields between Jackrong/Qwopus3.6-27B-v2 and Qwen/Qwen3.6-27B: architecture, model_type, language_model_only, tie_word_embeddings, text model_type, 64 text layers, full_attention_interval 4, hidden size, intermediate size, attention geometry, linear-attention geometry, mamba_ssm_dtype, and 262144 max positions." }, { "label": "Jackrong Qwopus3.6 Coder MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/tree/ab9175af03e34068fd773eeb64c80461bec2a526", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwopus3.6-27B-Coder-MTP-Q5_K_M.gguf 19.535698496 GB, BF16 54.657731136 GB, IQ4_XS 15.420447296 GB, Q2_K 10.864589376 GB, Q3_K_L 14.559795776 GB, Q3_K_M 13.500734016 GB, Q3_K_S 12.256533056 GB, Q4_K_M 16.810711616 GB, Q4_K_S 15.825295936 GB, Q5_K_S 18.971679296 GB, Q6_K 22.430997056 GB, Q8_0 29.047081536 GB, and mmproj-F32.gguf 0.931147296 GB. The API gguf.totalFileSize matches mmproj-F32, while the card's benchmark setup selects Q5_K_M." }, { "label": "Jackrong Qwopus3.6 Coder MTP Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/resolve/ab9175af03e34068fd773eeb64c80461bec2a526/Qwopus3.6-27B-Coder-MTP-Q5_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 57 metadata entries and 866 tensors. The linked file is 19.535698496 GB. Tensor spans sum to 19.524706304 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.874086400 GB, ordinary blk.0-63 tensors 17.303054336 GB, and blk.64 MTP tensors 0.304601088 GB. Metadata/tokenizer/header/file overhead accounts for 0.010992192 GB. Tensor spans split into Q5_K 14.987427840 GB, Q6_K 4.526592000 GB, and F32 0.010686464 GB. The header records qwen35.block_count 65, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, nextn_predict_layers 1, and no mmproj or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned Jackrong base config, Qwen3.6 base config comparison, linked-object HEAD checks, direct GGUF header/tensor-index range read of the selected Q5_K_M artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the Jackrong Coder MTP Q5_K_M GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer API-selected mmproj residency, other quantization bytes, multimodal projector execution, or MTP speculative acceleration unless a workload profile explicitly selects those paths." }, { "id": "jackrong--qwopus3-6-27b-v2-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwopus3.6-27B-v2-MTP-GGUF", "title": "Jackrong Qwopus3.6 27B v2 MTP GGUF Q5_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected Q5_K_M GGUF artifact of Jackrong's Qwopus3.6 27B v2 MTP package.", "model_family": "qwen3.6-dense-multimodal-v2-mtp", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "derived_package", "source": "Hugging Face API/card base_model metadata, Qwen3.6 base config, selected GGUF header metadata, linked-object HEAD checks, and Transformers qwen3_5 runtime review", "config_compatible": true, "notes": "The repo metadata and card identify the package as a Qwen/Qwen3.6-27B based MTP reasoning release. The target repo has no config.json, so high-level architecture fields are checked against the immutable Qwen/Qwen3.6-27B config and the selected GGUF header. The selected header records the same qwen35 text geometry plus one MTP block: 64 ordinary text layers, every fourth layer full attention, 4 KV heads, 256 full-attention key/value dimensions, 48 linear-attention layers, and 262144 context." }, "architecture": { "canonical_architecture_id": "qwopus3-6-27b-v2-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.320697856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.696097792, "resident_weight_gb": 19.535699136, "swept_weight_gb": 18.346018816, "auxiliary_resident_weight_gb": 1.18968032, "resident_parameter_scope": "selected Qwopus3.6-27B-v2-MTP-Q5_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.64 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; mmproj-F32.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The HF API gguf.totalFileSize points at mmproj-F32.gguf with architecture clip, so API GGUF metadata is not a valid language-model artifact selector for this repo. The card documents a GB10 local GGUF benchmark but does not name a quantization. This profile therefore audits Q5_K_M as the concrete practical main artifact matching Jackrong's sibling Qwopus3.6 MTP Q5_K_M benchmark convention; other quantizations should get separate profiles if selected. The selected linked file is 19.535699136 GB. Header tensor spans total 19.524706304 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010992832 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, ordinary blk.0-63 tensors, and blk.64 MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.64 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 base config and selected GGUF metadata record 64 ordinary text layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the selected Q5_K_M GGUF artifact. The MTP draft block is resident for the package but does not reduce token traffic in Bounds Engine v1; multimodal projector use requires a separate workload profile." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7150512493848942, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-v2-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected profile artifact is Q5_K_M because the API-selected GGUF block is the projector sidecar and the model card does not name a main quantization. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwopus3.6 v2 MTP GGUF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "api_artifact_caveat", "max_context_tokens" ], "notes": "The live HF API response at commit de55021a5b03540dc4a5ceec6bffff00e62fde65 records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-27B, 168444 downloads, region:us, text-generation-inference, qwen3_6, MTP, speculative-decoding, reasoning, coding, tool-use, and multimodal tags. The API gguf block reports architecture clip, gguf.total 460730096, and gguf.totalFileSize 931145760, which matches mmproj-F32.gguf rather than a language-model GGUF. This profile therefore does not use the API GGUF block as the text artifact selector." }, { "label": "Jackrong Qwopus3.6 v2 MTP GGUF model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "benchmark_context", "mtp_scope" ], "notes": "The pinned card records Apache-2.0 licensing, image-text-to-text GGUF packaging, base_model Qwen/Qwen3.6-27B, a speed-oriented MTP reasoning release, GB10 local GGUF server evaluation, llama-server total context 49152, and a 30-question benchmark reporting 10.46 overall tokens/sec versus 6.29 for Qwen3.6-27B. The card does not declare a serving-default GGUF quantization, so quant selection is recorded through linked-object and header evidence rather than card text." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The Qwen3.6 config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention key/value dimensions, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, mtp_num_hidden_layers 1, and resident vision config." }, { "label": "Jackrong Qwopus3.6 v2 MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF/tree/de55021a5b03540dc4a5ceec6bffff00e62fde65", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwopus3.6-27B-v2-MTP-Q5_K_M.gguf 19.535699136 GB, BF16 54.657731776 GB, IQ4_XS 15.420447936 GB, Q2_K 10.864590016 GB, Q3_K_L 14.559796416 GB, Q3_K_M 13.500734656 GB, Q3_K_S 12.256533696 GB, Q4_K_M 16.810712256 GB, Q4_K_S 15.825296576 GB, Q5_K_S 18.971679936 GB, Q6_K 22.430997696 GB, Q8_0 29.047082176 GB, and mmproj-F32.gguf 0.931145760 GB. The API gguf.totalFileSize matches mmproj-F32, while this profile audits Q5_K_M as the concrete selected main text artifact." }, { "label": "Jackrong Qwopus3.6 v2 MTP Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF/resolve/de55021a5b03540dc4a5ceec6bffff00e62fde65/Qwopus3.6-27B-v2-MTP-Q5_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 36 metadata entries and 866 tensors. The linked file is 19.535699136 GB. Tensor spans sum to 19.524706304 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.874086400 GB, ordinary blk.0-63 tensors 17.303054336 GB, and blk.64 MTP tensors 0.304601088 GB. Metadata/tokenizer/header/file overhead accounts for 0.010992832 GB. Tensor spans split into Q5_K 14.987427840 GB, Q6_K 4.526592000 GB, and F32 0.010686464 GB. The header records qwen35.block_count 65, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, nextn_predict_layers 1, and no mmproj or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned missing repo-config check, Qwen3.6 base config comparison, linked-object HEAD checks, direct GGUF header/tensor-index range read of the selected Q5_K_M artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the selected Jackrong Qwopus3.6 27B v2 MTP Q5_K_M GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer API-selected mmproj residency, other quantization bytes, multimodal projector execution, or MTP speculative acceleration unless a workload profile explicitly selects those paths." }, { "id": "jackrong--qwopus3-6-35b-a3b-v1-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jackrong/Qwopus3.6-35B-A3B-v1-GGUF", "title": "Jackrong Qwopus3.6 35B A3B v1 GGUF IQ4_XS", "summary": "Audited memory-side text-decode bounds profile for the API-selected IQ4_XS GGUF artifact of Qwopus3.6 35B A3B v1.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "unsloth/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card/API metadata, pinned Unsloth Qwen3.6 config, selected linked-object HEAD checks, selected GGUF header metadata, and existing audited Qwen3.6 runtime adapter evidence", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative and adapter merge of unsloth/Qwen3.6-35B-A3B. The selected IQ4_XS GGUF header records the same audited Qwen3.6 text geometry as the pinned Unsloth config: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 routed experts per token, and a separate always-on shared expert." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.939312512, "main_resident_weight_gb": 18.658150912, "auxiliary_resident_weight_gb": 0.2811616, "fixed_weight_gb": 1.335675392, "routed_expert_weight_gb": 0.06766592, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected IQ4_XS GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj-F32.gguf sidecar is not included unless explicitly loaded for another workload", "shared_expert_notes": "The model card and GGUF header record 8 routed experts plus one shared expert path. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected main GGUF mixes IQ4_XS-style packed tensors, Q8_K tensors, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_XS GGUF artifact after any multimodal prefill. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and speculative paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5464217778066173, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-xs-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative execution are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Jackrong Qwopus3.6 35B A3B v1 GGUF API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit bd297a4fe8dd8f9987c2c7a5b1029834c9a76c3f, the live API records a public Apache-2.0 image-text-to-text GGUF repo with base_model unsloth/Qwen3.6-35B-A3B, region:us, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 18939312512. Live downloads during audit were 97662, below the catalog inclusion threshold; the generated catalog row keeps its existing over-threshold snapshot count. The API totalFileSize matches Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf, so this profile targets that artifact." }, { "label": "Jackrong Qwopus3.6 35B A3B v1 GGUF model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF/raw/bd297a4fe8dd8f9987c2c7a5b1029834c9a76c3f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "layer_pattern", "routed_experts" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model unsloth/Qwen3.6-35B-A3B, Qwen3.6 35B/3B architecture, 256 routed experts, LoRA/SFT adaptation, and separate vision/mmproj sidecar guidance." }, { "label": "Unsloth Qwen3.6 35B A3B base config", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B/raw/2ab40a9acc6d567889ca4d4e59feb2da56121454/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The pinned base config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, and 262144 max position embeddings." }, { "label": "Jackrong Qwopus3.6 35B A3B v1 IQ4_XS linked object and GGUF range-read tensor index", "url": "https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF/resolve/bd297a4fe8dd8f9987c2c7a5b1029834c9a76c3f/Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 43 metadata entries and 733 tensors. The linked file is 18.939312512 GB. Tensor spans sum to 18.928323072 GB; metadata/tokenizer/header/file overhead accounts for 0.010989440 GB. token_embd.weight is 0.270172160 GB and resident-only; output.weight plus output_norm.weight is swept. Routed expert tensors sum to 17.322475520 GB, or 0.067665920 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 1.335675392 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found IQ4_XS 18.939312512 GB, Q4_K_M 21.166758272 GB, Q5_K_M 24.729131392 GB, Q6_K 28.514152832 GB, Q8_0 36.903139712 GB, and separate mmproj-F32 0.902822560 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned Unsloth Qwen3.6 config, selected linked file size, a direct GGUF header/tensor-index range read of the API-selected IQ4_XS artifact, sidecar HEAD checks, and the existing audited Qwen3.6 hybrid runtime adapter evidence." }, "notes": "Use this profile for the API-selected Jackrong Qwopus3.6 35B A3B v1 IQ4_XS main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "janhq--jan-v3-5-4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "janhq/Jan-v3.5-4B-gguf", "title": "JanHQ Jan v3.5 4B GGUF Q3_K_L", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q3_K_L GGUF artifact of Jan v3.5 4B.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "janhq/Jan-v3.5-4B", "relation": "quantized", "source": "Hugging Face model card/API metadata, Jan v3.5 source config and safetensors index, Jan v3 base config comparison, and direct GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo and model card identify this package as Jan-v3.5-4B, a fine-tuned Jan-v3-4B-base-instruct derivative. The Jan-v3.5 and Jan-v3 base configs differ only by transformers_version and record the same Qwen3 text geometry, but both configs set tie_word_embeddings true while the BF16 source index stores lm_head.weight and the selected GGUF stores output.weight. This profile therefore treats the config geometry as compatible but uses tensor evidence, not the tie flag, for output-projection traffic." }, "architecture": { "canonical_architecture_id": "jan-v3-5-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.411424256, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0.38895616, "resident_weight_gb": 2.406913376, "swept_weight_gb": 2.233828864, "auxiliary_resident_weight_gb": 0.173084512, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for Jan-v3.5-4B-Q3_K_L.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected main Q3_K_L GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as full matrices for each generated token", "notes": "The HF API gguf.totalFileSize matches Jan-v3.5-4B-Q3_K_L.gguf, so this profile targets the Q3_K_L artifact. Header tensor spans total 2.400958464 GB, while the linked file size is 2.406913376 GB. The main GGUF contains token_embd.weight, blk.* tensors, output.weight, and output_norm.weight. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Jan v3.5 config and selected GGUF metadata record full-context Qwen3 attention geometry with 36 layers, 8 KV heads, and 128 key/value head dimensions. The served config does not define sliding-window attention or recurrent state." }, "notes": "This profile models ordinary cached text decode for the selected main Q3_K_L GGUF artifact." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5456091358083152, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q3-k-l-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the JanHQ Q3_K_L GGUF. Tensor spans split into Q3_K 1.140339200 GB, Q5_K 0.940769280 GB, Q6_K 0.319065600 GB, and F32 0.000784384 GB. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters. Default GGUF KV is modeled as FP16 unless a workload profile explicitly chooses quantized KV." }, "evidence": [ { "label": "Jan v3.5 4B GGUF HF API metadata", "url": "https://huggingface.co/api/models/janhq/Jan-v3.5-4B-gguf", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 2b0fcab0066198e78f7c004a0815179fbce13985 records a public text-generation GGUF repo with Apache-2.0 license metadata, base_model janhq/Jan-v3-4B-base-instruct, endpoints_compatible, region:us, 304639 downloads, GGUF architecture qwen3, context length 262144, gguf.total 4411424256, and gguf.totalFileSize 2406913376." }, { "label": "Jan v3.5 4B GGUF model card", "url": "https://huggingface.co/janhq/Jan-v3.5-4B-gguf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model janhq/Jan-v3-4B-base-instruct, and says Jan-v3.5-4B is a fine-tuned variant of Jan-v3-4B-base-instruct. It lists the Qwen3-4B architecture, 4.0B parameters, 3.6B non-embedding parameters, 36 layers, GQA with 32 Q heads and 8 KV heads, and 262144 native context." }, { "label": "Jan v3.5 4B source HF API metadata", "url": "https://huggingface.co/api/models/janhq/Jan-v3.5-4B", "source_type": "derived_calculation", "supports": [ "base_model_proof", "total_params_b", "source_tensor_layout" ], "notes": "The live source repo API response at commit 53509d2d88feb0a1fccadf26e185383fc6a75d8e records BF16 safetensors parameters 4411424256, matching the selected GGUF logical tensor total. The source repo tags identify it as a fine-tune of janhq/Jan-v3-4B-base-instruct." }, { "label": "Jan v3.5 4B source config", "url": "https://huggingface.co/janhq/Jan-v3.5-4B/raw/53509d2d88feb0a1fccadf26e185383fc6a75d8e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter" ], "notes": "The immutable config records Qwen3ForCausalLM, bfloat16 dtype, 36 full-attention layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, no sliding-window attention, vocab size 151936, and rope_theta 5000000. It sets tie_word_embeddings true, but direct tensor evidence shows a separate output projection." }, { "label": "Jan v3.5 4B source safetensors index", "url": "https://huggingface.co/janhq/Jan-v3.5-4B/raw/53509d2d88feb0a1fccadf26e185383fc6a75d8e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "source_tensor_layout", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The source index records total_parameters 4411424256 and total_size 8822848512 across 399 BF16 tensors. Its weight_map includes both lm_head.weight and model.embed_tokens.weight, confirming that ordinary text decode should charge the output projection separately from resident-only input embeddings." }, { "label": "Jan v3 base config comparison", "url": "https://huggingface.co/janhq/Jan-v3-4B-base-instruct/raw/f331713ecaef85c9da95fe5b63131ddb0b8c6744/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "A structured comparison between Jan-v3.5-4B and Jan-v3-4B-base-instruct found the checked configs differ only by transformers_version. The base package API and safetensors index also record 4411424256 BF16 parameters and both lm_head.weight and model.embed_tokens.weight." }, { "label": "Jan v3.5 4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/janhq/Jan-v3.5-4B-gguf/tree/2b0fcab0066198e78f7c004a0815179fbce13985", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found Jan-v3.5-4B-Q3_K_L.gguf is 2406913376 bytes, exactly matching API gguf.totalFileSize. Sibling linked sizes range from 2054124896 bytes for Q3_K_S to 8829195616 bytes for the full GGUF. The model card llama.cpp example uses Q8_0, but the API-selected artifact is Q3_K_L, consistent with the selected-artifact rule used by existing GGUF profiles." }, { "label": "Jan v3.5 4B Q3_K_L GGUF range-read tensor index", "url": "https://huggingface.co/janhq/Jan-v3.5-4B-gguf/resolve/2b0fcab0066198e78f7c004a0815179fbce13985/Jan-v3.5-4B-Q3_K_L.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 28 metadata entries and 399 tensors. The linked file is 2.406913376 GB. Tensor spans sum to 2.400958464 GB: token_embd.weight 0.167129600 GB, blk.* tensors 1.914753024 GB, output.weight 0.319065600 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.005954912 GB. Tensor spans split into Q3_K 1.140339200 GB, Q5_K 0.940769280 GB, Q6_K 0.319065600 GB, and F32 0.000784384 GB. The header records qwen3.block_count 36, context_length 262144, embedding_length 2560, feed_forward_length 9728, attention.head_count 32, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 5000000, output.weight, and no mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, pinned Jan v3.5 source config and safetensors index, pinned Jan v3 base config comparison, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q3_K_L artifact." }, "notes": "Use this profile for the JanHQ main Q3_K_L GGUF text artifact. Do not infer sibling quantization sizes such as Q8_0 from this selected-artifact profile." }, { "id": "jeffcookio--mistral-small-3-2-24b-instruct-2506-awq-sym", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym", "title": "jeffcookio Mistral Small 3.2 24B Instruct 2506 AWQ Sym", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors symmetric INT4 package of Mistral Small 3.2 24B Instruct 2506.", "model_family": "mistral-small-3-2-24b-2506-multimodal-awq", "base_model_proof": { "base_model": "mistralai/Mistral-Small-3.2-24B-Instruct-2506", "relation": "quantized", "source": "Hugging Face API metadata, model card, served compressed-tensors config, base model config/params comparison, and direct safetensors shard headers", "config_compatible": true, "notes": "The repo API tags identify mistralai/Mistral-Small-3.2-24B-Instruct-2506 as the quantized base model. The model card also lists unsloth/Mistral-Small-3.2-24B-Instruct-2506 as a packaging source. Manual comparison against the pinned Mistral base config found matching checked text and vision geometry; the derivative adds compressed-tensors INT4 metadata, tokenizer/template changes, and unsloth_fixed true." }, "architecture": { "canonical_architecture_id": "mistral-small-3-2-24b-2506-mistral3", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 24.01136128, "swept_params_b": 22.90131456, "auxiliary_resident_params_b": 1.11004672, "resident_weight_gb": 15.02535104, "swept_weight_gb": 12.8052576, "auxiliary_resident_weight_gb": 2.22009344, "resident_parameter_scope": "true Mistral3 model parameters represented by the compressed checkpoint; excludes BF16 scale tensors and I64 weight-shape side tensors from logical parameter counts while charging their stored bytes", "swept_parameter_scope": "ordinary text decode charges language_model.model.layers.*, language_model.model.norm.weight, and language_model.lm_head.weight from the compressed-tensors package", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower.*, and multi_modal_projector.* are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary generated text token", "notes": "Direct safetensors header reads across all four shards matched index total_size 15.025351040 GB. Ordinary text/logit traffic is 12.805257600 GB: text layers 11.463070080 GB, final norm 0.000010240 GB, and lm_head 1.342177280 GB. Resident-only auxiliary bytes are 2.220093440 GB: input embedding 1.342177280 GB, vision tower 0.806610944 GB, and multimodal projector 0.071305216 GB. Stored text-layer traffic includes packed INT4 weights, BF16 scales, and tiny I64 weight-shape side tensors." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config and base config record 40 text layers, 8 KV heads, 128 head dimension, use_cache true, max_position_embeddings 131072, and sliding_window null. This profile charges full-context K and V streams for all text layers." }, "notes": "This profile models ordinary text decode after any image prefill. Vision tower, projector execution, image token processing, and multimodal prefill throughput are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6257600668611488, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-pack-quantized-int4-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed INT4 weights, BF16 scales, I64 weight-shape side tensors, and unquantized BF16 tensors from safetensors headers. Dequantization kernels, activation traffic, vision prefill, multimodal projector execution, scheduler behavior, and tool-calling tokenizer differences are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized INT4 weights with group_size 128, symmetric true, dynamic false, minmax observer, and ignored unquantized vision tower, projector, and lm_head modules. The model card says the checkpoint was created with llm-compressor and tested with vLLM main." }, "evidence": [ { "label": "jeffcookio Mistral Small 3.2 AWQ Sym API metadata", "url": "https://huggingface.co/api/models/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "tags", "safetensors_parameter_split", "commit_sha" ], "notes": "The current HF CLI/API response records commit 49ec31ec0975e46a1aeef24186a45f11be3042b8, public ungated status, 105570 downloads, tags safetensors, mistral3, dataset:nlphuji/flickr30k, base_model:mistralai/Mistral-Small-3.2-24B-Instruct-2506, base_model:quantized:mistralai/Mistral-Small-3.2-24B-Instruct-2506, compressed-tensors, and region:us. The package has no explicit license tag in the API response. Safetensors metadata reports BF16 1955220480, I64 560, I32 22229811200, and total 24185032240 entries; the direct audit treats BF16 scales and I64 shape tensors as compression side tensors for logical model-parameter counts." }, { "label": "jeffcookio Mistral Small 3.2 AWQ Sym served config", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym/raw/49ec31ec0975e46a1aeef24186a45f11be3042b8/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Mistral3ForConditionalGeneration, model_type mistral3, text_config MistralForCausalLM with hidden_size 5120, intermediate_size 32768, 40 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, sliding_window null, rope_theta 1000000000, vocab_size 131072, BF16 dtype, and a Pixtral vision tower with 24 layers, hidden size 1024, 16 heads, 64 head dimension, patch size 14, and image size 1540. The compressed-tensors config records pack-quantized INT4 Linear targets with group_size 128 and symmetric true, while ignoring all vision tower Linear modules, projector modules, and lm_head." }, { "label": "jeffcookio Mistral Small 3.2 AWQ Sym model card", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym/raw/49ec31ec0975e46a1aeef24186a45f11be3042b8/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format", "quantization" ], "notes": "The card lists base_model entries unsloth/Mistral-Small-3.2-24B-Instruct-2506 and mistralai/Mistral-Small-3.2-24B-Instruct-2506, says the checkpoint was created with llm-compressor, and says it worked with vLLM main on an RTX 3090 Ti as of 2025-07-01. Tool-calling tokenizer branch notes are runtime compatibility notes, not memory-bound geometry." }, { "label": "Mistral Small 3.2 24B Instruct base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mistral-Small-3.2-24B-Instruct-2506", "source_type": "model_card", "supports": [ "base_model_proof", "license", "base_parameter_count" ], "notes": "The base model API response records commit 95a6d26c4bfb886c58daf9d3f7332c857cb27b43, Apache-2.0 license tag, vLLM library, public ungated status, safetensors BF16 total 24011361280, and tags including mistral3, mistral-common, and region:us." }, { "label": "Mistral Small 3.2 24B Instruct base config and params", "url": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/raw/95a6d26c4bfb886c58daf9d3f7332c857cb27b43/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter", "vision_geometry" ], "notes": "The pinned base config records Mistral3ForConditionalGeneration with the same checked text_config and vision_config geometry as the quantized derivative. The pinned params.json separately records dim 5120, n_layers 40, hidden_dim 32768, n_heads 32, n_kv_heads 8, head_dim 128, vocab_size 131072, max_position_embeddings 131072, and matching vision_encoder fields." }, { "label": "jeffcookio Mistral Small 3.2 AWQ Sym safetensors index and shard headers", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym/raw/49ec31ec0975e46a1aeef24186a45f11be3042b8/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 15025351040 bytes across four shards, matching direct range-read tensor spans. Headers contain 1145 tensors totaling 15.025351040 GB: BF16 3.910440960 GB, I32 11.114905600 GB, and I64 0.000004480 GB. Linked shard sizes sum to 15.025500320 GB, or 149280 bytes of safetensors header/container overhead above tensor payload. Grouped tensor bytes are text layers 11.463070080 GB, final text norm 0.000010240 GB, lm_head 1.342177280 GB, input embedding 1.342177280 GB, vision tower 0.806610944 GB, and multimodal projector 0.071305216 GB. Packed INT4 qweight tensors store 11.114905600 GB and represent 22.229811200B logical text weights; BF16 scale tensors store 0.347340800 GB and are charged as compression side traffic, not true model parameters." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, model card, served compressed-tensors config, base model API/config/params metadata, safetensors index, linked-object HEAD checks, and direct range reads for all four safetensors shard headers." }, "notes": "This profile supersedes the scraped flat 0.5 byte/parameter estimate by using exact stored compressed-tensors byte traffic and by separating ordinary text/logit decode from resident-only input embedding, vision tower, and multimodal projector bytes." }, { "id": "jiunsong--supergemma4-26b-uncensored-gguf-v2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Jiunsong/supergemma4-26b-uncensored-gguf-v2", "title": "Jiunsong SuperGemma4 26B Uncensored Fast v2 GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the single Jiunsong SuperGemma4 26B Uncensored Fast v2 Q4_K_M GGUF artifact.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "finetune", "source": "Hugging Face API/model-card metadata, Google base profile/config geometry, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo metadata identifies this package as a GGUF derivative of google/gemma-4-26B-A4B-it, while the model card describes it as exported from the SuperGemma Fast MLX line rather than plain stock weights. The selected Q4_K_M GGUF header records the same Gemma 4 26B A4B text architecture: 30 layers, five full-attention layers, 25 sliding-attention layers, 128 experts, 8 routed experts per token, 1024-token sliding window, and 262144-token max context." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.796015232, "main_resident_weight_gb": 16.780193728, "auxiliary_resident_weight_gb": 0.015821504, "fixed_weight_gb": 1.65001312, "routed_expert_weight_gb": 0.118204536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "shared_expert_notes": "The GGUF header records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects the repo's only main GGUF artifact, supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf. A GGUF v3 range-read found 658 tensors and 49 metadata entries. Tensor spans total 16.780193728 GB, while the linked file is 16.796015232 GB. Routed expert tensors total 15.130180608 GB across 30 layers and 128 expert indexes, or 0.118204536 GB per expert index. Non-expert tensor spans total 1.650013120 GB, including token_embd.weight because the selected file has no separate output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. Gemma 4 full-attention layers use K=V behavior, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile targets the repo's single main Q4_K_M text GGUF artifact. No mmproj, MTP, vision, image, or draft tensors are present in that selected file." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6656331265199108, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and non-autoregressive serving effects are outside Bounds Engine v1.", "notes": "The selected artifact is the only main Q4_K_M GGUF in the repo and exactly matches HF API gguf.totalFileSize. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Jiunsong SuperGemma4 26B Uncensored Fast v2 GGUF API metadata", "url": "https://huggingface.co/api/models/Jiunsong/supergemma4-26b-uncensored-gguf-v2", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit 3ea8c452a2b136875c0c8b529612bed39c81e27a, the live API records a public non-gated Gemma-license text-generation GGUF repo with base_model google/gemma-4-26B-A4B-it, region:us, 86717 live downloads, GGUF architecture gemma4, context_length 262144, gguf.total 25233142046, and gguf.totalFileSize 16796015232. The catalog retains the earlier qualifying scrape count of 104781 downloads because the row was already in the over-100k audit set." }, { "label": "Jiunsong SuperGemma4 26B Uncensored Fast v2 GGUF model card", "url": "https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-gguf-v2/raw/3ea8c452a2b136875c0c8b529612bed39c81e27a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact" ], "notes": "The card records Gemma licensing, text-generation packaging, base_model google/gemma-4-26B-A4B-it, Q4_K_M GGUF format, the included file name, Apple Silicon llama.cpp validation, and notes that this artifact was exported from the SuperGemma Fast MLX line." }, { "label": "Google Gemma 4 26B A4B IT audited profile", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "kv_adapter" ], "notes": "The existing audited Google profile records 30 layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, attention_k_eq_v true, one shared expert, tied embeddings, and 262144 max positions." }, { "label": "Jiunsong SuperGemma4 Q4_K_M linked GGUF size", "url": "https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-gguf-v2/tree/3ea8c452a2b136875c0c8b529612bed39c81e27a", "source_type": "manual_review", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "Expanded tree and HEAD checks found a single main artifact, supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf, at 16.796015232 GB, exactly matching API gguf.totalFileSize. The repo contains no separate mmproj, imatrix, or MTP sidecar file." }, { "label": "Jiunsong SuperGemma4 Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-gguf-v2/resolve/3ea8c452a2b136875c0c8b529612bed39c81e27a/supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 658 tensors and 49 metadata entries. Tensor spans sum to 16.780193728 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015821504 GB. Tensor spans split into Q4_K 9.347272704 GB, Q8_0 3.863078912 GB, Q5_0 2.856730624 GB, Q6_K 0.667054080 GB, and F32 0.046057408 GB. Non-expert tensor spans total 1.650013120 GB. Routed expert tensors total 15.130180608 GB across 30 layers and 128 expert indexes, or 0.118204536 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, K=V full-attention geometry, separate sliding-layer K/V projections, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from live HF API metadata, pinned model card, the existing Google Gemma 4 base profile, selected linked-object HEAD checks, expanded linked-file metadata, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the Jiunsong Q4_K_M GGUF text artifact in ordinary text-decode bounds. Do not infer the upstream MLX line, other quantization levels, or speculative/multimodal sidecars unless a workload profile explicitly selects and audits them." }, { "id": "junhowie--qwen3-14b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "JunHowie/Qwen3-14B-GPTQ-Int4", "title": "JunHowie Qwen3 14B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the JunHowie GPTQ Int4 Qwen3 14B repo.", "model_family": "qwen3-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen3-14B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served GPTQ config, base Qwen3-14B config comparison, quantization config, safetensors index, and direct shard-header metadata", "config_compatible": true, "notes": "The API metadata and model card identify Qwen/Qwen3-14B as the base model. Manual comparison against the base config found matching Qwen3ForCausalLM geometry, context fields, no-sliding-attention settings, vocabulary size, and untied embedding setting; the GPTQ repo changes transformers_version, adds GPTQ quantization metadata, and stores quantized linear weights." }, "architecture": { "canonical_architecture_id": "qwen3-14b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7683072, "swept_params_b": 13.99039488, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 9.98427648, "swept_weight_gb": 8.42845184, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "GPTQ logical serving parameters reconstructed from qweight tensors plus BF16/F16 model tensors in indexed safetensors shards", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer qweight/qzeros/g_idx/scales tensors, layer norms, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full input embedding matrix for each generated token", "notes": "GPTQ qweight tensors are packed I32 values; logical parameter counts use the Hugging Face API safetensors logical counts. qzeros, g_idx, and scales are storage/serving metadata and are charged in stored-byte traffic. BF16 model weights and F16 scales are included in byte traffic. The config records tie_word_embeddings false and lm_head false in the GPTQ quantization config, so lm_head.weight is unquantized BF16 output traffic while model.embed_tokens.weight is resident-only for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records 40 full-attention layers, sliding_window null, and use_sliding_window false, so this profile charges full-context K and V streams for all 40 language layers." }, "notes": "Dense Qwen3ForCausalLM GPTQ profile using the served quantized repo config. The model card rounds native context to 32768 tokens and mentions 131072 with YaRN, but the pinned config and base config record 40960 max position embeddings with no rope_scaling." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6760609963476383, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-gptq-int4-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored GPTQ bytes from safetensors headers: packed I32 qweight/qzeros/g_idx tensors plus F16 scales and BF16 embeddings, output head, norms, and layernorm weights. Dequantization, activation traffic, compute, vLLM kernel behavior, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and GPTQ 4-bit quantization with group_size 128, desc_act false, symmetric quantization, true_sequential true, checkpoint_format gptq, pack_dtype int32, and lm_head false. KV cache is charged as FP16." }, "evidence": [ { "label": "JunHowie Qwen3 14B GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/JunHowie/Qwen3-14B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 5073a3e15216fb4facc39c953934910dbf0df51b, the API reports a public Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen3-14B, tags 4-bit, gptq, text-generation-inference, endpoints_compatible, region:us, current downloads 184023, and safetensors logical parameters I32 13212057600, BF16 1556249600, total 14768307200." }, { "label": "JunHowie Qwen3 14B GPTQ Int4 model card", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as Qwen/Qwen3-14B quantized to 4-bit with group size 128, recommends vLLM >= 0.9.2, records 14.8B total parameters, 13.2B non-embedding parameters, 40 layers, GQA with 40 Q heads and 8 KV heads, native 32768 context, and 131072 context with YaRN. The pinned config records 40960 max position embeddings with no rope_scaling, so the profile uses the config context." }, { "label": "JunHowie Qwen3 14B GPTQ Int4 served config", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "embedding_layout" ], "notes": "The config records Qwen3ForCausalLM with hidden_size 5120, intermediate_size 17408, 40 full-attention layers, 40 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, torch_dtype bfloat16, vocab_size 151936, rope_theta 1000000, and GPTQ quantization with bits 4, group_size 128, desc_act false, symmetric quantization, checkpoint_format gptq, pack_dtype int32, true_sequential true, and lm_head false." }, { "label": "JunHowie Qwen3 14B GPTQ Int4 quantize config", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/quantize_config.json", "source_type": "config", "supports": [ "quantization", "serving" ], "notes": "The quantize_config records bits 4, group_size 128, desc_act false, hyb_act false, symmetric quantization, lm_head false, quant_method gptq, checkpoint_format gptq, pack_dtype int32, and quantizer gptqmodel:4.0.0." }, { "label": "Qwen3 14B base config", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/40c069824f4251a91eefaf281ebe4c544efd3e18/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "embedding_layout" ], "notes": "The base config records matching Qwen3ForCausalLM geometry, hidden size, intermediate size, layer count, attention head count, KV head count, head dimension, context, no-sliding-attention fields, vocabulary size, RoPE theta, and untied embeddings. The base repo API reports BF16 safetensors total 14768307200 parameters." }, { "label": "JunHowie Qwen3 14B GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 9984276480 bytes across three shards. Expanded tree metadata found linked shard sizes 3994061392, 3998164048, and 1992195352 bytes. Direct range-read shard headers contain 1283 tensors totaling 9.984276480 GB: I32 6.665338880 GB, BF16 3.112499200 GB, and F16 0.206438400 GB. Safetensors header overhead across shards is 0.000144312 GB. Stored suffix bytes are qweight 6.606028800 GB, qzeros 0.051609600 GB, g_idx 0.007700480 GB, scales 0.206438400 GB, and BF16 weight tensors 3.112499200 GB. Tensor groups are model.embed_tokens.weight 1.555824640 GB, lm_head.weight 1.555824640 GB, model.layers tensors 6.872617600 GB, model.norm.weight 0.000010240 GB, layer norms 0.000819200 GB, MLP tensors 5.561057280 GB, and self-attention tensors 1.310740480 GB. Ordinary swept traffic is all tensor payload except resident-only model.embed_tokens.weight, totaling 8.428451840 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned GPTQ config, quantize_config, base Qwen3-14B config comparison, safetensors index, expanded linked-object sizes, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by using exact GPTQ stored bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "kaitchup--phi-3-mini-4k-instruct-gptq-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "kaitchup/Phi-3-mini-4k-instruct-gptq-4bit", "title": "Kaitchup Phi-3 Mini 4K Instruct GPTQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the GPTQ 4-bit Phi-3 Mini 4K Instruct repo.", "model_family": "phi3-gptq-dense", "base_model_proof": { "base_model": "microsoft/Phi-3-mini-4k-instruct", "relation": "quantized", "source": "Served config _name_or_path, quantization_config, and direct base config comparison", "config_compatible": false, "notes": "The served config records _name_or_path microsoft/Phi-3-mini-4k-instruct and GPTQ quantization metadata. Manual comparison against the pinned Microsoft BF16 base config found matching hidden size, layer count, attention geometry, context length, vocabulary size, and untied embeddings after excluding quantization metadata and dtype. The target config changes sliding_window from the base profile's 2047 to 2048 and references Microsoft remote code through auto_map." }, "architecture": { "canonical_architecture_id": "phi-3-mini-4k-gptq", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.82209536, "swept_params_b": 3.723594752, "auxiliary_resident_params_b": 0.098500608, "resident_weight_gb": 2.28137984, "swept_weight_gb": 2.084378624, "auxiliary_resident_weight_gb": 0.197001216, "resident_parameter_scope": "safetensors_header_stored_gptq_int4_i32_f16", "swept_parameter_scope": "ordinary text decode through model.layers.*, model.norm.weight, and lm_head.weight, excluding resident-only input embedding", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "The HF API records logical GPTQ parameters I32: 3623878656 and F16: 198216704, total 3822095360. Direct safetensors header parsing found one model.safetensors file with 707 tensors and 2.281379840 GB of stored payload: qweight 1.811939328 GB, qzeros 0.014155776 GB, g_idx 0.002228224 GB, scales 0.056623104 GB, biases 0.002031616 GB, layer norms 0.000393216 GB, final norm 0.000006144 GB, input embedding 0.197001216 GB, and lm_head 0.197001216 GB. The linked file is 2.281459480 GB, so safetensors header/file overhead is 0.000079640 GB and is not charged as accelerator-resident tensor payload." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 32, "kv_heads": 32, "head_dim": 96, "window_tokens": 2048, "kv_scalar_multiplier": 2, "notes": "The target config records sliding_window 2048. With the Microsoft Phi3 FlashAttention cache path, cached K/V are sliced to sliding_window - 1 before appending the current token, giving 2048 live K/V tokens." } ], "notes": "All 32 layers use the same FP16 sliding-window K/V geometry." }, "notes": "Phi3ForCausalLM is a dense text-only decoder model. This profile models ordinary autoregressive text decode for the GPTQ serving artifact." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.596892443834787, "kv_store_format": "fp16_sliding_window_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16_sliding_window_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-gptq-int4-phi3-sliding-window-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored GPTQ safetensors bytes, including packed qweight, qzeros, g_idx, scales, biases, embeddings, and lm_head tensors. GPTQ dequantization, activation traffic, kernel behavior, and cache writes are outside Bounds Engine v1.", "notes": "The served config records torch_dtype float16 plus GPTQ 4-bit quantization with group size 128, symmetric quantization, desc_act false, and true_sequential true. KV cache is charged as FP16 because no KV quantization is declared." }, "evidence": [ { "label": "Kaitchup Phi-3 Mini GPTQ API metadata", "url": "https://huggingface.co/api/models/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit", "source_type": "model_card", "supports": [ "repo", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At repo SHA 8d0c383a88f9b355bbc7ac0c36625fa7b9421e71, the HF API records a public non-gated Transformers text-generation repo with safetensors, phi3, custom_code, 4-bit, gptq, endpoints_compatible, region:us, current downloads 319947, and safetensors parameters I32: 3623878656 plus F16: 198216704." }, { "label": "Kaitchup Phi-3 Mini GPTQ served config", "url": "https://huggingface.co/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit/raw/8d0c383a88f9b355bbc7ac0c36625fa7b9421e71/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, _name_or_path microsoft/Phi-3-mini-4k-instruct, float16 dtype, tie_word_embeddings false, 32 layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 4096 max position embeddings, original 4096 context, sliding_window 2048, vocab size 32064, and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and dataset c4." }, { "label": "Microsoft Phi-3 Mini 4K Instruct base config", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_comparison" ], "notes": "The base config comparison confirms the same Phi3 text geometry after excluding quantization metadata and dtype. The target config sets sliding_window 2048 while the pinned Microsoft config records 2047." }, { "label": "Kaitchup Phi-3 Mini GPTQ safetensors header", "url": "https://huggingface.co/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit/resolve/8d0c383a88f9b355bbc7ac0c36625fa7b9421e71/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "lm_head_layout", "weight_format" ], "notes": "The linked model.safetensors file is 2.281459480 GB. A direct range read found a 79632-byte safetensors header and 2.281379840 GB of stored tensor payload across 707 tensors. Tensor grouping records model.layers.* 1.887371264 GB, model.norm.weight 0.000006144 GB, lm_head.weight 0.197001216 GB, and model.embed_tokens.weight 0.197001216 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served GPTQ config, pinned Microsoft base config comparison, linked-object HEAD check, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped ideal 4-bit estimate by using exact stored GPTQ bytes and the target repo's explicit 2048-token sliding-window cache setting." }, { "id": "largitdata--gemma-4-26b-a4b-it-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "LargitData/gemma-4-26b-a4b-it-fp8", "title": "LargitData Gemma 4 26B A4B IT FP8 Dynamic Norouter", "summary": "Audited memory-side text-decode bounds profile for the LargitData FP8 dynamic Gemma 4 26B A4B IT vLLM serving artifact.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison against the pinned Google BF16 base config found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, expert routing fields, and vision geometry. The LargitData artifact adds compressed-tensors FP8 dynamic quantization metadata while preserving the base architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 27.161975452, "main_resident_weight_gb": 26.01638662, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 3.14596102, "routed_expert_weight_gb": 0.1786752, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges one tied BF16 vocabulary projection, non-expert language tensors, and expected distinct routed expert tensors", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The base Gemma card describes 1 shared expert and the config records top_k_experts 8. The FP8 dynamic quantization recipe ignores router, embed, vision, audio, norm, and lm_head patterns; shared/always-on MLP tensors remain non-expert language traffic and are charged in fixed_weight_gb.", "notes": "Header-derived bytes are used because the artifact stores FP8 weights plus BF16 scale tensors and unquantized BF16 modules. The checkpoint stores model.language_model.embed_tokens.weight and has no separate lm_head.weight tensor; because config tie_word_embeddings is true, swept decode traffic charges one full BF16 vocabulary projection through the tied embedding tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Image/audio/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, not vision/audio encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-fp8-kv-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8 weights, BF16 scale tensors, unquantized BF16 modules, and FP8 KV bytes. Dynamic activation quantization, FP8 dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors float quantization with FP8 weights, dynamic per-token FP8 activation quantization, bfloat16 model dtype, and kv_cache_scheme null. The model card's primary serving target and launch command explicitly use --kv-cache-dtype fp8, so this profile represents that documented vLLM serving configuration rather than a BF16-KV launch." }, "evidence": [ { "label": "LargitData Gemma 4 26B FP8 Hugging Face API metadata", "url": "https://huggingface.co/api/models/LargitData/gemma-4-26b-a4b-it-fp8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 2398f3b38d1b2e56d14498f7da3a74bada8cddff, the public API reports a non-gated license:other text-generation repo with base_model google/gemma-4-26B-A4B-it, FP8/compressed-tensors/vLLM tags, region:us, 150515 downloads, and safetensors parameters split across BF16: 1339272526 and F8_E4M3: 24483430400 tensors, total 25822702926." }, { "label": "LargitData Gemma 4 26B FP8 config", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8/raw/2398f3b38d1b2e56d14498f7da3a74bada8cddff/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors float quantization, FP8 weight quantization with per-channel static scaling, dynamic per-token FP8 activation quantization, ignore patterns for routers, vision, audio, embeddings, norms, and lm_head, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 30 text layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 128 experts, 8 experts per token, and 262144 max position embeddings." }, { "label": "LargitData Gemma 4 26B FP8 model card", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "kv_store_format", "kv_read_format", "runtime_format", "license" ], "notes": "The card identifies the repo as an offline FP8_DYNAMIC checkpoint derived from google/gemma-4-26B-A4B-it for vLLM, documents the primary target as vllm/vllm-openai:gemma4, records test environment KV cache dtype fp8, and gives a launch command with --kv-cache-dtype fp8." }, { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The base card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this FP8 dynamic artifact; the LargitData artifact adds quantization_config while preserving the base architecture." }, { "label": "LargitData Gemma 4 26B FP8 safetensors index and shard headers", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8/raw/2398f3b38d1b2e56d14498f7da3a74bada8cddff/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "tie_word_embeddings" ], "notes": "The index records total_size 27161975452 bytes across two shards. Range-read safetensors headers found 1278 tensors with tensor payloads totaling 27.161975452 GB: 24.4834304 GB F8_E4M3 tensors and 2.678545052 GB BF16 tensors. model.language_model tensors total 26.01638662 GB, no separate lm_head.weight tensor is stored, model.language_model.embed_tokens.weight totals 1.476395008 GB, and resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic charges one tied vocabulary projection plus non-expert language tensors for 3.14596102 GB. Routed expert tensors total 22.8704256 GB and divide exactly into 128 uniform expert groups of 0.1786752 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from the current Hugging Face API metadata, model card, served config, base config comparison, safetensors index, and direct two-shard safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this FP8 Dynamic Norouter repo, including the catalog row's incorrect routed-experts-per-token value and BF16-KV legacy fallback display." }, { "id": "leon-se--gemma-4-e4b-it-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "leon-se/gemma-4-E4B-it-FP8-Dynamic", "title": "leon-se Gemma 4 E4B IT FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the leon-se FP8 dynamic compressed-tensors package of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, quantization recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo metadata records google/gemma-4-E4B-it as its quantized base model. Manual comparison found no differences across 24 checked text, vision, audio, context, tied-embedding, and attention geometry fields between the leon-se config and the pinned Google base config. The leon-se artifact adds compressed-tensors FP8 dynamic quantization metadata while preserving the base architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.66860081, "swept_params_b": 4.700851498, "auxiliary_resident_params_b": 3.967749312, "resident_weight_gb": 13.30935906, "swept_weight_gb": 5.373860436, "auxiliary_resident_weight_gb": 7.935498624, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.language_model.embed_tokens.weight input lookup and model.language_model.embed_tokens_per_layer.weight, and includes language-model layer tensors plus lm_head.weight output projection", "auxiliary_scope": "model.language_model.embed_tokens.weight, model.language_model.embed_tokens_per_layer.weight, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision are resident for token lookup, multimodal prefill, and PLE packaging but not swept as full matrices for each ordinary text decode token", "notes": "Header-derived bytes are used because the artifact stores FP8 language weights plus BF16 embeddings, lm_head, norms, activation range tensors, and unquantized audio/vision modules. The config records tie_word_embeddings true, but the safetensors file stores separate BF16 model.language_model.embed_tokens.weight and lm_head.weight tensors; resident bytes include both stored tensors, while swept decode traffic charges one full BF16 vocabulary projection rather than double-counting both. The large per-layer embedding table remains resident-only for ordinary decode, matching the audited Google Gemma 4 E4B PLE convention." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, with separate K and V projections." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 35 layers use 512-token local sliding-window attention with separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The compressed-tensors config has kv_cache_scheme null, so this profile charges BF16 KV cache bytes like the Google base profile. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-gemma4-e4b-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus BF16 embeddings, lm_head, norms, activation range tensors, and unquantized multimodal tensors from the safetensors header. Dynamic activation quantization, FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors float quantization with FP8 weights, dynamic per-token FP8 activation quantization, bfloat16 model dtype, and kv_cache_scheme null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "leon-se Gemma 4 E4B IT FP8 Dynamic API metadata", "url": "https://huggingface.co/api/models/leon-se/gemma-4-E4B-it-FP8-Dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 56e30bf603d18a4972caffafa1bb4a4f9a841dee, the API records a public non-gated repo with base_model google/gemma-4-E4B-it, base_model:quantized metadata, compressed-tensors and region:us tags, 182449 downloads, and safetensors parameters split across BF16: 4640757322 and F8_E4M3: 4027842560 tensors. The API does not record a license or pipeline tag for this repo." }, { "label": "leon-se Gemma 4 E4B IT FP8 Dynamic config", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/raw/56e30bf603d18a4972caffafa1bb4a4f9a841dee/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, bfloat16 top-level/text/vision/audio dtype, compressed-tensors float-quantized FP8 weights with dynamic token activation quantization, kv_cache_scheme null, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "leon-se Gemma 4 E4B IT FP8 Dynamic recipe", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/raw/56e30bf603d18a4972caffafa1bb4a4f9a841dee/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier targets [Linear], scheme FP8_DYNAMIC, and ignores lm_head plus model.embed_audio, model.embed_vision, model.audio_tower, and model.vision_tower modules." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no differences in 24 checked architecture fields between the leon-se config and the pinned Google base config after excluding quantization metadata and repository bookkeeping." }, { "label": "leon-se Gemma 4 E4B IT FP8 Dynamic safetensors header", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/resolve/56e30bf603d18a4972caffafa1bb4a4f9a841dee/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The linked object is 13.309692724 GB with a 333656-byte safetensors header. Range-reading the single header found 2510 tensors with payloads totaling 13.309359060 GB: 9.281516500 GB BF16 tensors and 4.027842560 GB F8_E4M3 tensors. Header tensor elements total 8.668600810B. model.language_model tensors total 11.011005012 GB. model.language_model.embed_tokens.weight is BF16 [262144, 2560] and contributes 0.671088640B elements / 1.342177280 GB resident-only for ordinary decode. lm_head.weight is a separate BF16 tensor of the same shape and remains in swept decode traffic. model.language_model.embed_tokens_per_layer.weight is BF16 [262144, 10752] and contributes 2.818572288B elements / 5.637144576 GB resident-only for ordinary decode. Audio/embed_audio tensors total 0.617514496 GB, and vision/embed_vision tensors total 0.338662272 GB. Ordinary text swept traffic, defined as language_model tensors excluding embed_tokens and embed_tokens_per_layer plus lm_head, totals 4.700851498B elements / 5.373860436 GB. Auxiliary resident tensors, defined as input embedding plus per-layer embedding plus audio plus vision, total 3.967749312B elements / 7.935498624 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served compressed-tensors config, quantization recipe, base Google config comparison, linked-object HEAD check, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the generated metadata estimate, which treated the artifact as a flat ideal 1-byte dense model and missed BF16 embeddings, the separate BF16 lm_head, activation range tensors, multimodal towers, per-layer embedding residency, and hybrid Gemma 4 KV traffic." }, { "id": "lgai-exaone--exaone-3-5-32b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ", "title": "EXAONE 3.5 32B Instruct AWQ", "summary": "Audited memory-side text-decode bounds profile for LG AI Research's AWQ 4-bit EXAONE 3.5 32B Instruct repo.", "model_family": "exaone-3.5-dense-awq", "base_model_proof": { "base_model": "LGAI-EXAONE/EXAONE-3.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, served AWQ config, base config comparison, custom runtime review, and safetensors header audit", "config_compatible": true, "notes": "The AWQ model card records EXAONE-3.5-32B-Instruct as the base model. Manual comparison found matching audited geometry fields between the pinned AWQ config and the pinned base config: architecture, 64 layers, hidden size 5120, intermediate size 27392, 40 attention heads, 8 KV heads, head_dim 128, 32768 max positions, Llama-3 RoPE scaling, and untied embeddings. The AWQ artifact adds quantization_config, changes dtype from float32 to float16, and records use_cache false, but the custom runtime supports DynamicCache when use_cache is enabled for serving." }, "architecture": { "canonical_architecture_id": "exaone-3.5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.0032, "swept_params_b": 31.478912, "auxiliary_resident_params_b": 0.524288, "resident_weight_gb": 18.18002432, "swept_weight_gb": 17.13144832, "auxiliary_resident_weight_gb": 1.048576, "resident_parameter_scope": "logical EXAONE 3.5 32B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes transformer.wte.weight input lookup and includes transformer.h.* tensors, transformer.ln_f.weight, and lm_head.weight", "auxiliary_scope": "transformer.wte.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, and unquantized F16 embedding/head/norm tensors. Logical parameter counts match the base model; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 64 layers, 8 KV heads, 128-dimensional key/value heads, 32768 max positions, and no sliding-window field. The custom ExaoneAttention implementation caches key_states and value_states through DynamicCache when use_cache is enabled. This serving profile charges ordinary full-context FP16 K/V cache streams." }, "notes": "Dense ExaoneForCausalLM AWQ profile using the served repo config, custom runtime cache implementation, and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-gemm-exaone-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, activation traffic, kernel behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert lm_head. KV cache is charged at FP16 two bytes per scalar for ordinary cached serving." }, "evidence": [ { "label": "EXAONE 3.5 32B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b", "downloads" ], "notes": "At commit 278e45157895e03158845c2c873ef6a37bd15a0f, the API records a public non-gated Transformers text-generation repo with EXAONE license metadata, exaone, custom_code, 4-bit, awq, and region:us tags, 157661 downloads, and safetensors parameters I32 30953963520 plus F16 1049236480. The model card states EXAONE 3.5 32B Instruct AWQ uses W4A16g128 group-wise weight-only quantization, 64 layers, GQA with 40 Q heads and 8 KV heads, vocab size 102400, and 32768 context." }, { "label": "EXAONE 3.5 32B Instruct AWQ config", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/raw/278e45157895e03158845c2c873ef6a37bd15a0f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The served config records ExaoneForCausalLM, model_type exaone, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert lm_head, 64 layers, hidden size 5120, intermediate size 27392, 40 attention heads, 8 KV heads, head_dim 128, 32768 max positions, Llama-3 RoPE scaling from 8192 original positions, tie_word_embeddings false, use_cache false, and vocab size 102400." }, { "label": "EXAONE 3.5 32B Instruct base config", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct/raw/3d27aab477c042df2e174d36eea47b77c8871ef0/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the base config after excluding quantization_config, dtype, use_cache, and repository bookkeeping. The base config records torch_dtype float32 and use_cache true." }, { "label": "EXAONE custom runtime cache implementation", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/raw/278e45157895e03158845c2c873ef6a37bd15a0f/modeling_exaone.py", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_kv_geometry", "embedding_layout" ], "notes": "Manual review found ExaoneAttention constructing q_proj, k_proj, and v_proj from hidden_size to attention/KV head dimensions, updating past_key_values with key_states and value_states when a cache is provided, and ExaoneModel creating DynamicCache when use_cache is enabled. The model uses transformer.wte for input embeddings, transformer.h decoder layers, transformer.ln_f final norm, and ExaoneForCausalLM's lm_head for logits." }, { "label": "EXAONE 3.5 32B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/resolve/278e45157895e03158845c2c873ef6a37bd15a0f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The index records total_size 18180024320 bytes across four shards. Range-read safetensors headers found 1475 tensors totaling 18.180024320 GB: 15.597895680 GB I32 tensors and 2.582128640 GB F16 tensors. AWQ tensor subclasses total 15.476981760 GB qweight, 0.120913920 GB qzeros, 0.483655680 GB scales, and 2.098472960 GB unquantized F16 weights. transformer.wte.weight is F16 with shape [102400, 5120] and contributes 0.524288000B logical parameters / 1.048576000 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. transformer.h.* tensors plus transformer.ln_f.weight plus lm_head.weight total 17.131448320 GB swept traffic." }, { "label": "EXAONE license", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/raw/278e45157895e03158845c2c873ef6a37bd15a0f/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records license other with license_name exaone. The LICENSE file is EXAONE AI Model License Agreement 1.1 - NC." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served AWQ config, base config comparison, custom runtime cache review, license file, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, and unquantized embedding/head tensors." }, { "id": "lilarest--gemma-4-31b-it-nvfp4-turbo", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "LilaRest/gemma-4-31B-it-NVFP4-turbo", "title": "LilaRest Gemma 4 31B IT NVFP4 Turbo", "summary": "Audited memory-side text-decode bounds profile for LilaRest's text-only ModelOpt NVFP4 Gemma 4 31B IT turbo artifact.", "model_family": "gemma4-dense-text", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct text-geometry comparison with audited Gemma 4 31B profiles", "config_compatible": false, "notes": "The repo records google/gemma-4-31B-it as its quantized base model and preserves the Gemma 4 31B text geometry, but it serves a Gemma4ForCausalLM language package rather than the full Gemma4ForConditionalGeneration multimodal wrapper. Manual checks found matching 60-layer text geometry, local/global attention pattern, tied embeddings, context fields, and ModelOpt NVFP4/FP8-KV metadata with the audited NVIDIA Gemma 4 31B NVFP4 text adapter." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-text", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.52776664, "swept_params_b": 32.52776664, "auxiliary_resident_params_b": 0, "resident_weight_gb": 19.29502292, "swept_weight_gb": 19.29502292, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_nvfp4_bf16_f8_e4m3_f32", "swept_parameter_scope": "all model.language_model tensors in the text-only package, including the tied embedding/output projection", "auxiliary_scope": "no resident-only multimodal tensors are present in this text-only package", "notes": "Header-derived bytes are used because this ModelOpt artifact stores NVFP4 packed weights, FP8 scale tensors, F32 scalar scales, and unquantized BF16 tensors. The config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config records attention_k_eq_v true, num_global_key_value_heads 4, and global_head_dim 512; full-attention layers use the K=V path." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. Sliding layers use separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. This text-only artifact declares FP8 KV cache through ModelOpt metadata." }, "notes": "The repo ships Gemma4ForCausalLM language weights without resident vision tensors. This profile models ordinary text decode and intentionally treats the tied embedding table as swept output-projection traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-text-only-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored tensor bytes, including ModelOpt payload and scale tensors, plus FP8 KV bytes. NVFP4 dequantization, activation traffic, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records ModelOpt NVFP4 weights and a static 8-bit float KV cache scheme. The artifact has no hf_quant_config sidecar; the config.json quantization_config is the authoritative serving metadata." }, "evidence": [ { "label": "LilaRest Gemma 4 31B NVFP4 Turbo API metadata", "url": "https://huggingface.co/api/models/LilaRest/gemma-4-31B-it-NVFP4-turbo", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "downloads" ], "notes": "At commit 77fb4a2380f1d167473fb11fa99e3d982bebbc2b, the API records an Apache-2.0 text-generation repo with base_model google/gemma-4-31B-it, NVFP4, ModelOpt, vLLM, lighthouse, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 182086. API safetensors metadata records BF16 1410617660, F8_E4M3 1830420480, U8 29286727680, and total 32527765820 storage elements." }, { "label": "LilaRest Gemma 4 31B NVFP4 Turbo config", "url": "https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo/raw/77fb4a2380f1d167473fb11fa99e3d982bebbc2b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "kv_store_format", "kv_read_format", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForCausalLM, ModelOpt NVFP4 quantization, static 8-bit float KV cache storage, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 32 attention heads, 16 sliding KV heads, 4 global KV heads, 256 sliding head dimension, 512 global head dimension, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison against the audited Google BF16 base and NVIDIA NVFP4 profiles found matching Gemma 4 31B text geometry. Unlike the full base repo, the LilaRest config is a text-only Gemma4ForCausalLM package and does not include resident vision tower fields." }, { "label": "LilaRest Gemma 4 31B NVFP4 Turbo safetensors index and shard headers", "url": "https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo/resolve/77fb4a2380f1d167473fb11fa99e3d982bebbc2b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "resident_weight_gb", "swept_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 19295022920 bytes across four shards. Range-read safetensors headers found 2062 tensors totaling 19.295022920 GB: 14.643363840 GB U8, 1.830420480 GB F8_E4M3, 2.821235320 GB BF16, and 0.000003280 GB F32. The tensor names are all under model.language_model; no vision tensors and no separate lm_head.weight are present. model.language_model.embed_tokens.weight contributes 2.818572288 GB and is swept because it is the tied output projection." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned config, base config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds on the text-only turbo artifact. It deliberately differs from full multimodal Gemma 4 profiles by having no resident-only vision package and no separate lm_head tensor." }, { "id": "llmfan46--gemma-4-31b-it-uncensored-heretic-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF", "title": "llmfan46 Gemma 4 31B IT Uncensored Heretic GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected BF16 GGUF artifact of the Heretic/ARA decensored Gemma 4 31B IT derivative.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "derived_package", "source": "Hugging Face model card/API metadata, uncensored base config, Google Gemma 4 31B IT config/profile, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF card describes this as a decensored version of google/gemma-4-31B-it made with Heretic v1.2.0 and ARA, and the base full-precision repo records base_model google/gemma-4-31B-it. The accessible base derivative config and selected GGUF header preserve the same Gemma 4 31B text geometry: dense 60-layer text model, 50 sliding layers, ten full-attention layers, 1024-token sliding window, 262144 context, tied embeddings, and no MoE." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 61.413188736, "swept_weight_gb": 61.397356416, "auxiliary_resident_weight_gb": 0.01583232, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31B-it-uncensored-heretic-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and tensor-alignment/file overhead are resident in the selected artifact file but not swept as model tensors; gemma-4-31B-it-mmproj-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor payloads total 61.397354736 GB, with 0.000001680 GB of tensor-alignment padding, while the linked file size is 61.413188736 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config and GGUF metadata record four global KV heads and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads, 256 local key/value length, and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the full-precision BF16 GGUF because HF API gguf.totalFileSize matches gemma-4-31B-it-uncensored-heretic-BF16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant BF16 tensor format." }, "evidence": [ { "label": "llmfan46 Gemma 4 31B Heretic GGUF HF API metadata", "url": "https://huggingface.co/api/models/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF CLI/API response at commit eee61b81461ac75eb920a24ca9e5d420bb66e33d records base_model llmfan46/gemma-4-31B-it-uncensored-heretic, Apache-2.0 license metadata, image-text-to-text pipeline, region:us, 131249 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, and gguf.totalFileSize 61413188736. The API totalFileSize selects the BF16 main file." }, { "label": "llmfan46 Gemma 4 31B Heretic GGUF model card", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF/raw/eee61b81461ac75eb920a24ca9e5d420bb66e33d/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "runtime_format" ], "notes": "The card records Apache-2.0 licensing, base_model llmfan46/gemma-4-31B-it-uncensored-heretic, says this is a decensored version of google/gemma-4-31B-it made using Heretic v1.2.0 with the ARA method, and lists gemma-4-31B-it-uncensored-heretic-BF16.gguf as the full-precision quantization. It also lists gemma-4-31B-it-mmproj-BF16.gguf as a separate vision projector required for multimodal capability." }, { "label": "llmfan46 Gemma 4 31B Heretic base config", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic/raw/031ce2063177d066ff637f15ab59a2ca2eccfd19/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The accessible base derivative config records Gemma4ForConditionalGeneration, bfloat16 dtype, tie_word_embeddings true, dense Gemma 4 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, four global KV heads, 512 global key/value length, 16 local KV heads, 256 local key/value length, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT audited profile and config", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "tie_word_embeddings" ], "notes": "The existing audited Google Gemma 4 31B IT profile and immutable config establish the same layered local/global attention rule, K=V full-attention behavior, tied embedding/output projection rule, and multimodal sidecar exclusion pattern used here." }, { "label": "llmfan46 Gemma 4 31B Heretic GGUF linked-object HEAD checks", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF/tree/eee61b81461ac75eb920a24ca9e5d420bb66e33d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-31B-it-uncensored-heretic-BF16.gguf is 61413188736 bytes, exactly matching API gguf.totalFileSize. Sibling main quantizations are Q8_0 32635675776 bytes, Q6_K 25201484928 bytes, Q5_K_M 21845570688 bytes, Q5_K_S 21311841408 bytes, Q4_K_M 18687063168 bytes, Q4_K_S 17763165312 bytes, Q3_K_L 16628270208 bytes, and Q3_K_M 15287108736 bytes. The sidecar gemma-4-31B-it-mmproj-BF16.gguf is 1200726208 bytes and is not the selected main text artifact." }, { "label": "llmfan46 Gemma 4 31B Heretic BF16 GGUF range-read tensor index", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF/resolve/eee61b81461ac75eb920a24ca9e5d420bb66e33d/gemma-4-31B-it-uncensored-heretic-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 51 metadata entries and 833 tensors. The selected file is 61.413188736 GB, with tensor payloads starting at byte 15832320. Tensor spans total 61.397356416 GB across 30.697345596B logical elements: token_embd.weight 2.818572288 GB, blk.* tensors 58.578761600 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Tensor payloads split into BF16 61.392027648 GB and F32 0.005327088 GB, with 1680 bytes of tensor-alignment padding. Metadata/tokenizer/header/file overhead accounts for 0.015832320 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, a layer KV head array with ten global layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF CLI/API metadata, pinned model card, accessible base derivative config, existing Google Gemma 4 31B IT profile/config, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the llmfan46 main BF16 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "llmfan46--gemma-4-e4b-it-ultra-uncensored-heretic-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF", "title": "llmfan46 Gemma 4 E4B IT Ultra Uncensored Heretic GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected BF16 GGUF artifact of the Heretic/ARA decensored Gemma 4 E4B IT derivative.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "derived_package", "source": "Hugging Face model card/API metadata, uncensored source config, Google Gemma 4 E4B IT config/profile convention, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF card describes this as a decensored version of google/gemma-4-E4B-it made with Heretic v1.2.0 and ARA, and the base full-precision repo records base_model google/gemma-4-E4B-it. The accessible base derivative config and selected GGUF header preserve the same Gemma 4 E4B text geometry: dense 42-layer text model, seven full-attention layers, 35 sliding-window layers, 512-token sliding window, 131072 context, tied embeddings, per-layer embeddings, and no MoE." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.51806929, "swept_params_b": 4.699497002, "auxiliary_resident_params_b": 2.818572288, "resident_weight_gb": 15.053094752, "swept_weight_gb": 9.400126784, "auxiliary_resident_weight_gb": 5.652967968, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges GGUF tensor spans in the selected main artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; gemma-4-E4B-it-mmproj-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "Gemma 4 E4B uses per-layer embeddings. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per-layer projection tensors inside blk.*, normal language blocks, token_embd.weight, output_norm.weight, per_layer_proj_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The source config records 2 KV heads and the selected GGUF header records 512 key/value length for global attention with separate K and V streams." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the source config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2.0022520541384323, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the full-precision BF16 GGUF because HF API gguf.totalFileSize exactly matches gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "llmfan46 Gemma 4 E4B Heretic GGUF API metadata", "url": "https://huggingface.co/api/models/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 1465f37b7dbd15e91241ae78ffebecb9f25e15de records base_model llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic, Apache-2.0 license metadata, any-to-any pipeline, region:us, 92138 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 7518069290, and gguf.totalFileSize 15053094752. The API totalFileSize selects the BF16 main file." }, { "label": "llmfan46 Gemma 4 E4B Heretic GGUF model card", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF/raw/1465f37b7dbd15e91241ae78ffebecb9f25e15de/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "runtime_format" ], "notes": "The card records Apache-2.0 licensing, base_model llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic, says this is a decensored version of google/gemma-4-E4B-it made using Heretic v1.2.0 with the ARA method, and lists gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf as the full-precision quantization. It also lists gemma-4-E4B-it-mmproj-BF16.gguf as a separate vision projector required for multimodal capability." }, { "label": "llmfan46 Gemma 4 E4B Heretic source API metadata", "url": "https://huggingface.co/api/models/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "revision" ], "notes": "The source model API records commit 5964fe4c7339c5974e879baba8982a09616f68ca, base_model google/gemma-4-E4B-it, base_model:finetune metadata, Apache-2.0 license, any-to-any pipeline, region:us, and BF16 safetensors total 7996156490." }, { "label": "llmfan46 Gemma 4 E4B Heretic source config", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic/raw/5964fe4c7339c5974e879baba8982a09616f68ca/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "tie_word_embeddings", "per_layer_embeddings", "max_context_tokens" ], "notes": "The accessible source config records Gemma4ForConditionalGeneration, tie_word_embeddings true, dense Gemma 4 text config, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, 2 KV heads, 256 local head dimension, 512 global key/value length in the GGUF header, 131072 max position embeddings, 262144 vocab size, and resident vision metadata." }, { "label": "Google Gemma 4 E4B IT audited profile and config", "url": "https://huggingface.co/google/gemma-4-E4B-it", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The existing audited Google Gemma 4 E4B IT profile and immutable config establish the same layered local/global attention rule, tied embedding/output projection rule, and per-layer embedding residency convention used here." }, { "label": "llmfan46 Gemma 4 E4B Heretic GGUF linked-object HEAD checks", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF/tree/1465f37b7dbd15e91241ae78ffebecb9f25e15de", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf is 15053094752 bytes, exactly matching API gguf.totalFileSize. Sibling main quantizations are Q8_0 8031240032 bytes, Q6_K 6217260896 bytes, Q5_K_M 5762912096 bytes, Q5_K_S 5685968736 bytes, and Q4_K_M 5335289696 bytes. The sidecar gemma-4-E4B-it-mmproj-BF16.gguf is 991552000 bytes and is not the selected main text artifact." }, { "label": "llmfan46 Gemma 4 E4B Heretic BF16 GGUF range-read tensor index", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF/resolve/1465f37b7dbd15e91241ae78ffebecb9f25e15de/gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 720 tensors. The selected file is 15.053094752 GB, with tensor payloads starting at byte 15823392. Tensor spans total 15.037271360 GB across 7518069290 logical elements. per_layer_token_embd.weight is 5.637144576 GB / 2818572288 logical elements and is resident-only for this ordinary text-decode profile. Swept tensor spans excluding that lookup table total 9.400126784 GB / 4699497002 logical elements. token_embd.weight is 1.342177280 GB and no output.weight tensor is stored, so token_embd.weight remains swept as tied output-projection traffic. Stored tensor spans split into BF16 15.035006976 GB and F32 0.002264384 GB. GGUF metadata/tokenizer/header/file overhead accounts for 0.015823392 GB. The header records gemma4.block_count 42, context_length 131072, attention.head_count 8, attention.head_count_kv 2, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, embedding_length_per_layer_input 256, and no mmproj, vision, or audio tensors in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from live HF API metadata, pinned model card, accessible source derivative config, existing Google Gemma 4 E4B IT profile/config convention, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the llmfan46 main BF16 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload. The repo currently has fewer than 100k live downloads, but it remains profiled because it was already present in the over-100k catalog queue." }, { "id": "llmfan46--qwen3-6-35b-a3b-uncensored-heretic-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF", "title": "llmfan46 Qwen3.6 35B A3B Uncensored Heretic GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of llmfan46's Heretic derivative of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "llmfan46/Qwen3.6-35B-A3B-uncensored-heretic", "relation": "quantized", "source": "GGUF repo API metadata and model card, source Heretic repo API/config metadata, Qwen base config, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF quantization of llmfan46/Qwen3.6-35B-A3B-uncensored-heretic. The source repo records Qwen/Qwen3.6-35B-A3B as its base model and keeps the same memory-relevant text geometry as the selected GGUF header and Qwen base config: 40 text blocks, full_attention_interval 4, 10 full-attention layers, 30 DeltaNet linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 256 routed experts, 8 experts per token, one shared expert, and 262144 context length." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 69.376641152, "main_resident_weight_gb": 68.348529152, "auxiliary_resident_weight_gb": 1.028112, "fixed_weight_gb": 3.924019712, "routed_expert_weight_gb": 0.25165824, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected BF16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj BF16 sidecar is not included unless explicitly selected by another workload", "shared_expert_notes": "The Qwen/source config records shared_expert_intermediate_size 512, and the selected GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected BF16 GGUF stores BF16 tensors plus tiny F32 tensors. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The source config, Qwen base config, and selected GGUF metadata record 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 main GGUF artifact. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and multimodal prefill require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-moe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected BF16 GGUF artifact. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, image prefill, and write traffic are outside Bounds Engine v1.", "notes": "The API-selected artifact is Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf because HF API gguf.totalFileSize exactly matches that linked object. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param records nominal BF16 tensor payload size." }, "evidence": [ { "label": "llmfan46 Qwen3.6 35B A3B Heretic GGUF API metadata", "url": "https://huggingface.co/api/models/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit f70be2db155a4192a59c559ece01572f3cd508ab records a public Apache-2.0 image-text-to-text GGUF repo with 145539 downloads, base_model llmfan46/Qwen3.6-35B-A3B-uncensored-heretic, region:us, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 69376641152. The API totalFileSize matches Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf, so this profile targets that artifact." }, { "label": "llmfan46 Qwen3.6 35B A3B Heretic GGUF model card", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF/raw/f70be2db155a4192a59c559ece01572f3cd508ab/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "runtime_format" ], "notes": "The card frontmatter records Apache-2.0 licensing, image-text-to-text packaging, and base_model llmfan46/Qwen3.6-35B-A3B-uncensored-heretic. The card describes these as GGUF quantizations of that source model and says the source is a Heretic v1.2.0 derivative of Qwen/Qwen3.6-35B-A3B. It lists BF16 as the full-precision quantization and a separate mmproj BF16 sidecar for multimodal use." }, { "label": "llmfan46 Qwen3.6 35B A3B Heretic source API metadata", "url": "https://huggingface.co/api/models/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family" ], "notes": "The source Heretic repo is public, Apache-2.0 licensed, image-text-to-text, and records base_model Qwen/Qwen3.6-35B-A3B at commit dbfd9eb0cdc7c33fc970b06429f6e043b6851190. Its API safetensors block reports 35107181936 BF16 parameters, while the pinned config gives the memory-relevant architecture used for compatibility checks." }, { "label": "llmfan46 Qwen3.6 35B A3B Heretic source config", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic/raw/dbfd9eb0cdc7c33fc970b06429f6e043b6851190/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, and a 27-layer Qwen3.5 vision config." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "linear_attention_state" ], "notes": "The public Qwen base config records the same audited text decode geometry as the llmfan46 source config and the selected GGUF header, with bfloat16 text dtype. This confirms the Heretic text stack preserves the Qwen3.6 35B A3B memory-relevant layout." }, { "label": "llmfan46 Qwen3.6 35B A3B Heretic BF16 GGUF linked object and range-read tensor index", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF/resolve/f70be2db155a4192a59c559ece01572f3cd508ab/Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 41 metadata entries and 733 tensors. The linked file is 69.376641152 GB. Tensor spans sum to 69.365647872 GB; metadata/tokenizer/header/file overhead accounts for 0.010993280 GB. Tensor spans split into BF16 69.276794880 GB and F32 0.088852992 GB. token_embd.weight is 1.017118720 GB and resident-only; output.weight is a separate 1.017118720 GB swept tensor. Routed expert tensors sum to 64.424509440 GB, or 0.251658240 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 3.924019712 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found BF16 69.376641152 GB, Q4_K_M 21.233608512 GB, Q5_K_M 24.764032832 GB, Q6_K 28.576057152 GB, Q8_0 36.903143232 GB, and separate mmproj BF16 0.902822240 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, source Heretic config/API metadata, Qwen base config, selected linked file sizes, a direct GGUF header/tensor-index range read of the API-selected BF16 artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected llmfan46 Qwen3.6 35B A3B Uncensored Heretic BF16 main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "lmg-anon--vntl-llama3-8b-v2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmg-anon/vntl-llama3-8b-v2-gguf", "title": "VNTL Llama 3 8B V2 GGUF Q5_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected Q5_K_M GGUF artifact of the VNTL Llama 3 8B V2 translation model.", "model_family": "llama-3-dense-gguf", "base_model_proof": { "base_model": "rinna/llama-3-youko-8b", "relation": "quantized", "source": "Hugging Face model card metadata, GGUF API metadata, base-model config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of rinna/llama-3-youko-8b. The card describes the model as a LLaMA 3 Youko QLoRA fine-tune for Japanese visual-novel translation. The selected GGUF header records the same Llama 3 8B text geometry as the public rinna base config." }, "architecture": { "canonical_architecture_id": "llama-3-8b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 5.732991008, "swept_weight_gb": 5.363982336, "auxiliary_resident_weight_gb": 0.369008672, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for vntl-llama3-8b-v2-hf-q5_k_m.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q5_K_M linked file is 5.732991008 GB. Header tensor spans total 5.725151232 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007839776 GB. The main GGUF contains output.weight, token_embd.weight, blk.* tensors, and output_norm.weight. Because output.weight is stored separately and the base config records untied embeddings, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The rinna base config and selected GGUF header record a Llama-style 32-layer decoder with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected Q5_K_M GGUF artifact. The repository also contains a Q8_0 artifact with a different linked size and tensor traffic, which requires its own profile if selected by a workload." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7139233495582533, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact uses a mixed Q5_K_M layout: tensor spans split into 4.467392512 GB Q5_K, 1.256693760 GB Q6_K, and 0.001064960 GB F32. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "VNTL Llama 3 8B V2 GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmg-anon/vntl-llama3-8b-v2-gguf", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit ab7c8285d386fab7cd89e951f62e129085df6372 records a public non-gated GGUF repo with base_model rinna/llama-3-youko-8b, translation pipeline, 750593 downloads, GGUF architecture llama, 8192 context length, gguf.total 8030261248, and gguf.totalFileSize 5732991008." }, { "label": "VNTL Llama 3 8B V2 GGUF model card", "url": "https://huggingface.co/lmg-anon/vntl-llama3-8b-v2-gguf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "translation_scope" ], "notes": "The card records the Llama 3 license, base_model rinna/llama-3-youko-8b, Japanese/English translation metadata, dataset lmg-anon/VNTL-v5-1k, and describes this repo as a LLaMA 3 Youko QLoRA fine-tune for Japanese visual-novel translation using the default LLaMA 3 prompt format." }, { "label": "Rinna Llama 3 Youko 8B API metadata", "url": "https://huggingface.co/api/models/rinna/llama-3-youko-8b", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "The live base-model API response at commit 6af890bc6294b8f311e25a1462fcd38554d95b8e records a public non-gated Transformers Llama text-generation repo, BF16 safetensors total 8030261248 parameters, Llama 3 license, and base_model meta-llama/Meta-Llama-3-8B." }, { "label": "Rinna Llama 3 Youko 8B config", "url": "https://huggingface.co/rinna/llama-3-youko-8b/raw/6af890bc6294b8f311e25a1462fcd38554d95b8e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable base config records LlamaForCausalLM, model_type llama, BF16 dtype, hidden size 4096, 32 layers, 32 attention heads, 8 KV heads, 128 head dimension, 14336 feed-forward size, 128256 vocab size, untied embeddings, and 8192 max position embeddings." }, { "label": "VNTL Llama 3 8B V2 GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmg-anon/vntl-llama3-8b-v2-gguf/tree/ab7c8285d386fab7cd89e951f62e129085df6372", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found vntl-llama3-8b-v2-hf-q5_k_m.gguf resolves to x-repo-commit ab7c8285d386fab7cd89e951f62e129085df6372, x-linked-size 5732991008, x-xet-hash e2738cb5180599c37403f56101239e9c9680aff071539db6f94db4d74285d9c3, and final content-length 5732991008. The repo also contains vntl-llama3-8b-v2-hf-q8_0.gguf at linked size 8540773792 bytes." }, { "label": "VNTL Llama 3 8B V2 Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmg-anon/vntl-llama3-8b-v2-gguf/resolve/ab7c8285d386fab7cd89e951f62e129085df6372/vntl-llama3-8b-v2-hf-q5_k_m.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 291 tensors. The linked file is 5.732991008 GB. Tensor spans sum to 5.725151232 GB: output.weight 0.430940160 GB, output_norm.weight 0.000016384 GB, token_embd.weight 0.361168896 GB, and blk.* tensors 4.933025792 GB. Metadata/tokenizer/header/file overhead accounts for 0.007839776 GB. Stored tensor bytes split into Q5_K 4.467392512 GB, Q6_K 1.256693760 GB, and F32 0.001064960 GB. The header records llama.block_count 32, context_length 8192, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 500000, and separate output.weight/token_embd.weight tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, public base-model API/config metadata, linked GGUF object checks, and a direct GGUF header/tensor-index range read of the selected Q5_K_M artifact." }, "notes": "Use this profile for the selected VNTL Llama 3 8B V2 Q5_K_M GGUF artifact. Do not infer the Q8_0 artifact's resident or traffic bytes from this profile." }, { "id": "lmstudio-community--deepseek-r1-0528-qwen3-8b-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit", "title": "LM Studio DeepSeek R1 0528 Qwen3 8B MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized DeepSeek R1 0528 Qwen3 8B artifact.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records deepseek-ai/DeepSeek-R1-0528-Qwen3-8B as its base model. Manual comparison found no differences across the checked text architecture fields between the MLX config and the audited BF16 DeepSeek source config. The MLX repo adds quantization and quantization_config fields plus packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "deepseek-r1-0528-qwen3-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 4.607731712, "swept_weight_gb": 4.257671168, "auxiliary_resident_weight_gb": 0.350060544, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "model tensors excluding model.embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as eight 4-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from the single-file safetensors header and are authoritative for the bound." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served MLX config matches the BF16 DeepSeek source config: use_sliding_window false, sliding_window null, 36 layers, 8 KV heads, 128 head dimension, BF16 runtime dtype, and 131072 max positions. Bounds Engine v1 charges full-context BF16 K and V streams for ordinary cached text decode." }, "notes": "Dense Qwen3ForCausalLM MLX profile using the served LM Studio config and tensor headers rather than deriving structure from the model name." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5625540942883057, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records 4-bit affine MLX quantization with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 1b6065ad0cd7256f662c9ef6fd0ed2dfd06e286f, the API records a public MIT text-generation repo with mlx, qwen3, conversational, 4-bit, and region:us tags, base_model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, 318926 downloads, and safetensors parameters BF16 256259072, U32 1023803392, total 1280062464 storage elements. The card describes this as a 4-bit MLX quantized version optimized for Apple Silicon." }, { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit/raw/1b6065ad0cd7256f662c9ef6fd0ed2dfd06e286f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM, qwen3, tie_word_embeddings false, 4-bit affine MLX quantization with group_size 64, bfloat16 runtime dtype, 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 head dimension, 131072 max position embeddings with YaRN factor 4 from 32768 original positions, use_sliding_window false, sliding_window null, use_cache true, and vocab size 151936." }, { "label": "DeepSeek R1 0528 Qwen3 8B BF16 source config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no differences in 18 checked text architecture fields: architectures, model_type, hidden size, intermediate size, layer count, attention/KV heads, head dimension, max positions, RoPE scaling, torch dtype, untied embeddings, vocab size, norm epsilon, activation, attention dropout, and token IDs. The source repo already has an audited BF16 profile." }, { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 4-bit safetensors index and header", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit/raw/1b6065ad0cd7256f662c9ef6fd0ed2dfd06e286f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index records total_size 4607731712 bytes for model.safetensors. Direct safetensors header range reads found 907 tensors totaling 4.607731712 GB: BF16 0.512518144 GB and U32 4.095213568 GB. The linked object is 4.607835164 GB including safetensors header/container overhead; bounds use tensor spans. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors excluding MLX .scales/.biases gives 8.190735360B resident parameters. model.embed_tokens contributes 0.622329856B logical params and 0.350060544 GB resident-only. lm_head contributes 0.622329856B logical params and 0.350060544 GB swept. Other model-body tensors contribute 6.946075648B logical params and 3.907610624 GB swept. Ordinary text swept traffic is model excluding embed_tokens plus lm_head, totaling 7.568405504B logical params and 4.257671168 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned BF16 DeepSeek source config comparison, model card, safetensors index, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and missed MLX scales, biases, BF16 side tensors, untied lm_head storage, and resident-only embedding bytes. It is an ordinary cached text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--deepseek-r1-0528-qwen3-8b-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit", "title": "LM Studio DeepSeek R1 0528 Qwen3 8B MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized DeepSeek R1 0528 Qwen3 8B artifact.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records deepseek-ai/DeepSeek-R1-0528-Qwen3-8B as its base model. Manual comparison found no differences across the checked text architecture fields between the MLX config and the audited BF16 DeepSeek source config. The MLX repo adds quantization and quantization_config fields plus packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "deepseek-r1-0528-qwen3-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 8.70294528, "swept_weight_gb": 8.041719808, "auxiliary_resident_weight_gb": 0.661225472, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "model tensors excluding model.embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as four 8-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from the two safetensors shard headers and are authoritative for the bound." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served MLX config matches the BF16 DeepSeek source config: use_sliding_window false, sliding_window null, 36 layers, 8 KV heads, 128 head dimension, BF16 runtime dtype, and 131072 max positions. Bounds Engine v1 charges full-context BF16 K and V streams for ordinary cached text decode." }, "notes": "Dense Qwen3ForCausalLM MLX profile using the served LM Studio config and tensor headers rather than deriving structure from the model name." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.062535279, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records 8-bit affine MLX quantization with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit e5242815680c2ed87df14b5a1deb76bd672dd69f, the API records a public MIT text-generation repo with mlx, qwen3, conversational, 8-bit, and region:us tags, base_model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, 291320 downloads, and safetensors parameters BF16 256259072, U32 2047606784, total 2303865856 storage elements. The card describes this as an 8-bit MLX quantized version optimized for Apple Silicon." }, { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit/raw/e5242815680c2ed87df14b5a1deb76bd672dd69f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM, qwen3, tie_word_embeddings false, 8-bit affine MLX quantization with group_size 64, bfloat16 runtime dtype, 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 head dimension, 131072 max position embeddings with YaRN factor 4 from 32768 original positions, use_sliding_window false, sliding_window null, use_cache true, and vocab size 151936." }, { "label": "DeepSeek R1 0528 Qwen3 8B BF16 source config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no differences in 18 checked text architecture fields: architectures, model_type, hidden size, intermediate size, layer count, attention/KV heads, head dimension, max positions, RoPE scaling, torch dtype, untied embeddings, vocab size, norm epsilon, activation, attention dropout, and token IDs. The source repo already has an audited BF16 profile." }, { "label": "LM Studio DeepSeek R1 Qwen3 8B MLX 8-bit safetensors index and headers", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit/raw/e5242815680c2ed87df14b5a1deb76bd672dd69f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index records total_size 8702945280 bytes across two safetensors shards and 907 tensors. Direct safetensors header range reads found 907 tensors totaling 8.702945280 GB: BF16 0.512518144 GB and U32 8.190427136 GB. Linked shard sizes total 8.703048738 GB including safetensors header/container overhead; bounds use tensor spans. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors excluding MLX .scales/.biases gives 8.190735360B resident parameters. model.embed_tokens contributes 0.622329856B logical params and 0.661225472 GB resident-only. lm_head contributes 0.622329856B logical params and 0.661225472 GB swept. Other model-body tensors contribute 6.946075648B logical params and 7.380494336 GB swept. Ordinary text swept traffic is model excluding embed_tokens plus lm_head, totaling 7.568405504B logical params and 8.041719808 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned BF16 DeepSeek source config comparison, model card, safetensors index, and direct two-shard safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit dense weights and missed MLX scales, biases, BF16 side tensors, untied lm_head storage, and resident-only embedding bytes. It is an ordinary cached text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-12b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-12B-it-GGUF", "title": "LM Studio Gemma 4 12B IT GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected LM Studio Q4_K_M GGUF artifact of Gemma 4 12B IT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google Gemma 4 12B IT config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it. The selected GGUF header records the same Gemma 4 12B text geometry as the Google config. The LM Studio GGUF repo does not ship config.json, so the immutable Google config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.381384864, "swept_weight_gb": 7.365559808, "auxiliary_resident_weight_gb": 0.015825056, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-12B-it-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; mmproj-gemma-4-12B-it-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_K_M GGUF file selected by HF API gguf.totalFileSize and the rendered usage examples. Header tensor spans total 7.365559808 GB, while the linked file size is 7.381384864 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6199049820159312, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_K_M GGUF because HF API gguf.totalFileSize matches gemma-4-12B-it-Q4_K_M.gguf and the model card usage examples select Q4_K_M. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 12B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-12B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit ad4c92b806b5efd34055a9343e022736ad182634 records base_model google/gemma-4-12B-it, Apache-2.0 license, region:us, 409335 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 7381384864." }, { "label": "LM Studio Gemma 4 12B GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-12B-it, the Q4_K_M artifact in llama.cpp usage examples, and says LM Studio produced the GGUF quantization using llama.cpp release b9512." }, { "label": "Google Gemma 4 12B IT config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "LM Studio Gemma 4 12B GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-GGUF/tree/ad4c92b806b5efd34055a9343e022736ad182634", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-12B-it-Q4_K_M.gguf is 7381384864 bytes, exactly matching API gguf.totalFileSize. The sibling quantizations are gemma-4-12B-it-Q6_K.gguf at 9786023584 bytes and gemma-4-12B-it-Q8_0.gguf at 12669648544 bytes. mmproj-gemma-4-12B-it-BF16.gguf is 175115296 bytes and is not the selected main text artifact." }, { "label": "LM Studio Gemma 4 12B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-GGUF/resolve/ad4c92b806b5efd34055a9343e022736ad182634/gemma-4-12B-it-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 45 metadata entries and 667 tensors. The selected file is 7.381384864 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15825056. Tensor spans total 7.365559808 GB across 11907350576 logical elements: token_embd.weight 0.8257536 GB, blk.* tensors 6.539789824 GB, output_norm.weight 0.00001536 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_K 5.24648448 GB, Q6_K 2.1159936 GB, and F32 0.003081728 GB, with 1344 bytes of tensor-alignment padding. Metadata/tokenizer/header/file overhead accounts for 0.015825056 GB. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google Gemma 4 12B IT config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-12b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-12B-it-QAT-GGUF", "title": "LM Studio Gemma 4 12B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for the selected LM Studio Q4_0 GGUF artifact of Gemma 4 12B IT QAT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 12B text geometry as the Google QAT unquantized config. The LM Studio GGUF repo does not ship config.json, so the immutable Google QAT unquantized config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 6.97587856, "swept_weight_gb": 6.960055808, "auxiliary_resident_weight_gb": 0.015822752, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-12B-it-QAT-Q4_0.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; mmproj-gemma-4-12B-it-QAT-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_0 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 6.960055808 GB, while the linked file size is 6.97587856 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5858464076853765, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-12B-it-QAT-Q4_0.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 12B QAT GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 291406f49e16eff811c85ad8884d375f34138663 records base_model google/gemma-4-12B-it-qat-q4_0-unquantized, Apache-2.0 license, region:us, 711351 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 6975878560." }, { "label": "LM Studio Gemma 4 12B QAT GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-12B-it-qat-q4_0-unquantized, and says LM Studio produced the GGUF quantization using llama.cpp release b9518." }, { "label": "Google Gemma 4 12B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "LM Studio Gemma 4 12B QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF/tree/291406f49e16eff811c85ad8884d375f34138663", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-12B-it-QAT-Q4_0.gguf is 6975878560 bytes, exactly matching API gguf.totalFileSize. mmproj-gemma-4-12B-it-QAT-BF16.gguf is 175115328 bytes and is not the selected main text artifact." }, { "label": "LM Studio Gemma 4 12B QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF/resolve/291406f49e16eff811c85ad8884d375f34138663/gemma-4-12B-it-QAT-Q4_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 45 metadata entries and 667 tensors. The selected file is 6.97587856 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15822752. Tensor spans total 6.960055808 GB across 11907350576 logical elements: token_embd.weight 0.8257536 GB, blk.* tensors 6.134285824 GB, output_norm.weight 0.00001536 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_0 6.13122048 GB, Q6_K 0.8257536 GB, and F32 0.003081728 GB, with 1344 bytes of tensor-alignment padding. Metadata/tokenizer/header/file overhead accounts for 0.015822752 GB. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for the LM Studio main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit", "title": "LM Studio Gemma 4 26B A4B IT MLX 4-bit", "summary": "Audited memory-side bounds profile for the LM Studio MLX 4-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model metadata, non-QAT Google config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio card metadata lists google/gemma-4-26B-A4B-it as the original model. Manual comparison found no differences in the audited text, vision, context, tied-embedding, expert, and attention geometry fields between this MLX config and the pinned non-QAT google/gemma-4-26B-A4B-it config. The MLX repo adds quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 15.608614044, "main_resident_weight_gb": 14.467688508, "auxiliary_resident_weight_gb": 1.140925536, "fixed_weight_gb": 1.621321788, "routed_expert_weight_gb": 0.10036224, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "vision_tower and embed_vision tensors are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The Gemma 4 26B config records top_k_experts 8, and the public Gemma card/profile evidence describes 1 shared expert. The MLX headers store dense language_model.model.layers.*.mlp.* and router tensors outside language_model.model.layers.*.experts.*, so those shared/always-on tensors are charged in fixed_weight_gb.", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. Header-derived bytes are authoritative for the bound: routed expert tensors total 12.846366720 GB across 30 layers and 128 stacked expert indexes, or 0.100362240 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma 4 26B profiles. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, not vision prefill or image encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5880250448923184, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-moe-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, and resident vision tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, and vision prefill are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 4-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the Google base logical total parameter count; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 3af5252ed1e675e6bba9be8cc3087bc00920799c, the API records a public/non-gated Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 4-bit, and region:us tags, plus 234475 downloads. The model card says this is a 4-bit quantized version of gemma-4-26B-A4B-it using MLX, optimized for Apple Silicon. The API safetensors block reports BF16 1358865998, U32 3222720512, and total 4581586510 storage elements." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, text_config.attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, 16 attention heads, 8 local KV heads, 2 global KV heads, 256 head dimension, 512 global head dimension, 262144 max position embeddings, resident vision config, and an MLX quantization_config with default 4-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit." }, { "label": "Google Gemma 4 26B A4B IT non-QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of audited architecture fields found no differences between the MLX config and the non-QAT Google instruction-tuned config: model type, layer count and types, sliding window, attention heads, KV heads, head dimensions, expert count, top_k_experts, tied embeddings, text_config.attention_k_eq_v, max position embeddings, vocabulary size, and vision geometry." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 15608614044 bytes across three shards. Range-read safetensors headers found 1697 tensors totaling the same 15.608614044 GB: BF16 2.717731996 GB and U32 12.890882048 GB. Stored-byte buckets are vision/embed_vision auxiliary tensors 1.140925536 GB, fixed language tensors 1.621321788 GB, and routed expert tensors 12.846366720 GB. Expert tensors are stacked with first dimension 128; all 30 layers have uniform expert bytes of 428212224 bytes, so the exact per-expert-index byte charge is 12.846366720 GB / 128 = 0.100362240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, non-QAT Google config comparison, model card text, safetensors index, and direct shard header range reads." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights with one routed expert per token. It uses exact MLX package bytes and the correct Gemma 4 26B routing geometry for ordinary text decode." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit", "title": "LM Studio Gemma 4 26B A4B IT MLX 5-bit", "summary": "Audited memory-side bounds profile for the LM Studio MLX 5-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model metadata, non-QAT Google config comparison, MLX sibling config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio card metadata lists google/gemma-4-26B-A4B-it as the original model. Manual comparison found no differences in the audited text, vision, context, tied-embedding, expert, and attention geometry fields between this MLX 5-bit config, the audited MLX 6-bit sibling config, and the pinned non-QAT google/gemma-4-26B-A4B-it config. The MLX repo adds 5-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.694814876, "main_resident_weight_gb": 17.553483836, "auxiliary_resident_weight_gb": 1.14133104, "fixed_weight_gb": 1.852368956, "routed_expert_weight_gb": 0.12266496, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "vision_tower and embed_vision tensors are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The Gemma 4 26B config records top_k_experts 8, and the public Gemma card/profile evidence describes 1 shared expert. The MLX headers store dense language_model.model.layers.*.mlp.* and router tensors outside language_model.model.layers.*.experts.*, so those shared/always-on tensors are charged in fixed_weight_gb.", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. Header-derived bytes are authoritative for the bound: routed expert tensors total 15.701114880 GB across 30 layers and 128 stacked expert indexes, or 0.122664960 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma 4 26B profiles. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, not vision prefill or image encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.7042918305061963, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-5bit-affine-moe-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, and resident vision tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, and vision prefill are outside Bounds Engine v1.", "notes": "The config records bfloat16 text and vision dtypes plus an MLX quantization_config with default 5-bit affine group_size 64 and 120 language MLP/router projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the same logical total parameter denominator used by the audited MLX siblings; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT MLX 5-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 13c9cdfe77d1aa897e7e94ab7d418ddf2345f219, the API records a public/non-gated Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 5-bit, and region:us tags, plus 199805 downloads. The model card says this is a 5-bit quantized version of gemma-4-26B-A4B-it using MLX, optimized for Apple Silicon. The API safetensors block reports BF16 1358865998, U32 3994270720, and total 5353136718 storage elements." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit/raw/13c9cdfe77d1aa897e7e94ab7d418ddf2345f219/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, text_config dtype bfloat16, text_config.attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, 16 attention heads, 8 local KV heads, 2 global KV heads, 256 head dimension, 512 global head dimension, 262144 max position embeddings, resident vision config, and an MLX quantization_config with default 5-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit." }, { "label": "Google Gemma 4 26B A4B IT non-QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of audited architecture fields found no differences between the MLX 5-bit config and the non-QAT Google instruction-tuned config: model type, layer count and types, sliding window, attention heads, KV heads, head dimensions, expert count, top_k_experts, tied embeddings, text_config.attention_k_eq_v, max position embeddings, vocabulary size, and vision geometry." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 6-bit sibling config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit/raw/fd6c729deddbd6211b37be13f2a42ef2603b3b56/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison", "quantization_comparison" ], "notes": "Manual comparison found no audited architecture differences between the MLX 5-bit config and the audited MLX 6-bit sibling config. The quantization metadata differs only in default bits: this repo records default 5-bit affine storage, while the 6-bit sibling records default 6-bit affine storage with the same 120 8-bit language MLP/router projection overrides." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 5-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit/raw/13c9cdfe77d1aa897e7e94ab7d418ddf2345f219/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 18694814876 bytes across four shards. Range-read safetensors headers found 1697 tensors totaling the same 18.694814876 GB: BF16 2.717731996 GB and U32 15.977082880 GB. Stored-byte buckets are vision/embed_vision auxiliary tensors 1.141331040 GB, fixed language tensors 1.852368956 GB, and routed expert tensors 15.701114880 GB. Expert tensors are stacked with first dimension 128; all 30 layers have uniform expert bytes of 523370496 bytes, so the exact per-expert-index byte charge is 15.701114880 GB / 128 = 0.122664960 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, non-QAT Google config comparison, audited MLX 6-bit sibling config comparison, model card text, safetensors index, and direct shard header range reads." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense bytes and one routed expert per token. It uses exact MLX package bytes and the correct Gemma 4 26B routing geometry for ordinary text decode." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit", "title": "LM Studio Gemma 4 26B A4B IT MLX 6-bit", "summary": "Audited memory-side bounds profile for the LM Studio MLX 6-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model metadata, non-QAT Google config comparison, MLX sibling config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio card metadata lists google/gemma-4-26B-A4B-it as the original model. Manual comparison found no differences in the audited text, vision, context, tied-embedding, expert, and attention geometry fields between this MLX 6-bit config, the audited MLX 4-bit sibling config, and the pinned non-QAT google/gemma-4-26B-A4B-it config. The MLX repo adds 6-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 21.781015708, "main_resident_weight_gb": 20.639279164, "auxiliary_resident_weight_gb": 1.141736544, "fixed_weight_gb": 2.083416124, "routed_expert_weight_gb": 0.14496768, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "vision_tower and embed_vision tensors are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The Gemma 4 26B config records top_k_experts 8, and the public Gemma card/profile evidence describes 1 shared expert. The MLX headers store dense language_model.model.layers.*.mlp.* and router tensors outside language_model.model.layers.*.experts.*, so those shared/always-on tensors are charged in fixed_weight_gb.", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. Header-derived bytes are authoritative for the bound: routed expert tensors total 18.555863040 GB across 30 layers and 128 stacked expert indexes, or 0.144967680 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma 4 26B profiles. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, not vision prefill or image encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8205586161200742, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-moe-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, and resident vision tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, and vision prefill are outside Bounds Engine v1.", "notes": "The config records bfloat16 text and vision dtypes plus an MLX quantization_config with default 6-bit affine group_size 64 and 120 language MLP/router projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the same logical total parameter denominator used by the audited MLX 4-bit and 8-bit siblings; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit fd6c729deddbd6211b37be13f2a42ef2603b3b56, the API records a public/non-gated Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 6-bit, and region:us tags, plus 200715 downloads. The model card says this is a 6-bit quantized version of gemma-4-26B-A4B-it using MLX, optimized for Apple Silicon. The API safetensors block reports BF16 1358865998, U32 4765820928, and total 6124686926 storage elements." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit/raw/fd6c729deddbd6211b37be13f2a42ef2603b3b56/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, text_config dtype bfloat16, text_config.attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, 16 attention heads, 8 local KV heads, 2 global KV heads, 256 head dimension, 512 global head dimension, 262144 max position embeddings, resident vision config, and an MLX quantization_config with default 6-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit." }, { "label": "Google Gemma 4 26B A4B IT non-QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of audited architecture fields found no differences between the MLX 6-bit config and the non-QAT Google instruction-tuned config: model type, layer count and types, sliding window, attention heads, KV heads, head dimensions, expert count, top_k_experts, tied embeddings, text_config.attention_k_eq_v, max position embeddings, vocabulary size, and vision geometry." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 4-bit sibling config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison", "quantization_comparison" ], "notes": "Manual comparison found no audited architecture differences between the MLX 6-bit config and the audited MLX 4-bit sibling config. The quantization metadata differs only in default bits: this repo records default 6-bit affine storage, while the 4-bit sibling records default 4-bit affine storage with the same 120 8-bit language MLP/router projection overrides." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit/raw/fd6c729deddbd6211b37be13f2a42ef2603b3b56/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 21781015708 bytes across five shards. Range-read safetensors headers found 1697 tensors totaling the same 21.781015708 GB: BF16 2.717731996 GB and U32 19.063283712 GB. Stored-byte buckets are vision/embed_vision auxiliary tensors 1.141736544 GB, fixed language tensors 2.083416124 GB, and routed expert tensors 18.555863040 GB. Expert tensors are stacked with first dimension 128; all 30 layers have uniform expert bytes of 618528768 bytes, so the exact per-expert-index byte charge is 18.555863040 GB / 128 = 0.144967680 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, non-QAT Google config comparison, audited MLX 4-bit sibling config comparison, model card text, safetensors index, and direct shard header range reads." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense bytes and one routed expert per token. It uses exact MLX package bytes and the correct Gemma 4 26B routing geometry for ordinary text decode." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit", "title": "LM Studio Gemma 4 26B A4B IT MLX 8-bit", "summary": "Audited memory-side bounds profile for the LM Studio MLX 8-bit package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model metadata, non-QAT Google config comparison, MLX sibling config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio card metadata lists google/gemma-4-26B-A4B-it as the original model. Manual comparison found no differences in the audited text, vision, context, tied-embedding, expert, and attention geometry fields between this MLX 8-bit config, the audited MLX 4-bit sibling config, and the pinned non-QAT google/gemma-4-26B-A4B-it config. The MLX repo adds 8-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 27.953417372, "main_resident_weight_gb": 26.81086982, "auxiliary_resident_weight_gb": 1.142547552, "fixed_weight_gb": 2.54551046, "routed_expert_weight_gb": 0.18957312, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "vision_tower and embed_vision tensors are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The Gemma 4 26B config records top_k_experts 8, and the public Gemma card/profile evidence describes 1 shared expert. The MLX headers store dense language_model.model.layers.*.mlp.* and router tensors outside language_model.model.layers.*.experts.*, so those shared/always-on tensors are charged in fixed_weight_gb.", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. Header-derived bytes are authoritative for the bound: routed expert tensors total 24.265359360 GB across 30 layers and 128 stacked expert indexes, or 0.189573120 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma 4 26B profiles. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, not vision prefill or image encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.0530921873478298, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-moe-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, and resident vision tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, and vision prefill are outside Bounds Engine v1.", "notes": "The config records bfloat16 text and vision dtypes plus an MLX quantization_config with default 8-bit affine group_size 64 and 120 language MLP/router projection overrides also at 8-bit. weight_bytes_per_param records resident stored bytes divided by the same logical total parameter denominator used by the audited MLX 4-bit sibling; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 669237cd8dad363224c976432475b81dd5db5a89, the API records a public/non-gated Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 8-bit, and region:us tags, plus 208960 downloads. The model card says this is an 8-bit quantized version of gemma-4-26B-A4B-it using MLX, optimized for Apple Silicon. The API safetensors block reports BF16 1358865998, U32 6308921344, and total 7667787342 storage elements." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit/raw/669237cd8dad363224c976432475b81dd5db5a89/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, text_config dtype bfloat16, text_config.attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, 16 attention heads, 8 local KV heads, 2 global KV heads, 256 head dimension, 512 global head dimension, 262144 max position embeddings, resident vision config, and an MLX quantization_config with default 8-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit." }, { "label": "Google Gemma 4 26B A4B IT non-QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of audited architecture fields found no differences between the MLX 8-bit config and the non-QAT Google instruction-tuned config: model type, layer count and types, sliding window, attention heads, KV heads, head dimensions, expert count, top_k_experts, tied embeddings, text_config.attention_k_eq_v, max position embeddings, vocabulary size, and vision geometry." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 4-bit sibling config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison", "quantization_comparison" ], "notes": "Manual comparison found no audited architecture differences between the MLX 8-bit config and the audited MLX 4-bit sibling config. The quantization metadata differs only in default bits: this repo records default 8-bit affine storage, while the 4-bit sibling records default 4-bit affine storage with the same 120 8-bit language MLP/router projection overrides." }, { "label": "LM Studio Gemma 4 26B A4B IT MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit/raw/669237cd8dad363224c976432475b81dd5db5a89/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 27953417372 bytes across six shards. Range-read safetensors headers found 1697 tensors totaling the same 27.953417372 GB: BF16 2.717731996 GB and U32 25.235685376 GB. Stored-byte buckets are vision/embed_vision auxiliary tensors 1.142547552 GB, fixed language tensors 2.545510460 GB, and routed expert tensors 24.265359360 GB. Expert tensors are stacked with first dimension 128; all 30 layers have uniform expert bytes of 808845312 bytes, so the exact per-expert-index byte charge is 24.265359360 GB / 128 = 0.189573120 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, non-QAT Google config comparison, audited MLX 4-bit sibling config comparison, model card text, safetensors index, and direct shard header range reads." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit MoE weights with one routed expert per token. It uses exact MLX package bytes and the correct Gemma 4 26B routing geometry for ordinary text decode." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF", "title": "LM Studio Gemma 4 26B A4B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for the LM Studio Q4_0 GGUF package of Gemma 4 26B A4B IT QAT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, existing Google QAT GGUF profile comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The LM Studio card and API metadata identify this package as a GGUF derivative of google/gemma-4-26B-A4B-it-qat-q4_0-unquantized. The repo does not include config.json, so architecture proof comes from the base QAT config, the direct LM Studio GGUF header, and comparison with the already audited Google official Q4_0 QAT GGUF profile." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 14.439362752, "main_resident_weight_gb": 14.423538808, "auxiliary_resident_weight_gb": 0.015823944, "fixed_weight_gb": 1.577172088, "routed_expert_weight_gb": 0.10036224, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for gemma-4-26B-A4B-it-QAT-Q4_0.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors; mmproj-gemma-4-26B-A4B-it-QAT-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for a multimodal workload", "shared_expert_notes": "The model card lineage and GGUF header record 8 active / 128 total routed experts and the Gemma 4 shared expert path. The GGUF header stores dense blk.*.ffn_down/gate/up tensors and tiny ffn_down_exps.scale tensors outside the routed Q4_0 expert weight span, so those shared/always-on bytes are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B logical elements and gguf.totalFileSize selects gemma-4-26B-A4B-it-QAT-Q4_0.gguf. A GGUF v3 range-read found 658 tensors and 46 metadata entries. Tensor spans total 14.423538808 GB, while the linked file is 14.439362752 GB. Routed expert Q4_0 weight tensors total 12.846366720 GB across 30 layers and 128 expert indexes, or 0.100362240 GB per expert index. Non-expert and always-on tensor spans total 1.577172088 GB. Tensor payloads match the official Google Q4_0 QAT GGUF main artifact; the LM Studio linked file has 1,312 additional resident overhead bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The base config and GGUF head-count/sliding-window arrays show five full-attention layers. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile targets the selected LM Studio main Q4_0 QAT text GGUF artifact. The multimodal projector sidecar is a separate workload concern and is not included unless explicitly loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5722380005501119, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is LM Studio's main Q4_0 GGUF because HF API gguf.totalFileSize matches gemma-4-26B-A4B-it-QAT-Q4_0.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT QAT GGUF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 43abd4da26c7b9c1a93c4d96a588c8a5374be8eb records a public non-gated Apache-2.0 GGUF repo, base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, region:us, 302235 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 25233142046, and gguf.totalFileSize 14439362752." }, { "label": "LM Studio Gemma 4 26B A4B IT QAT GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF/raw/43abd4da26c7b9c1a93c4d96a588c8a5374be8eb/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, and says the GGUF quantization was provided by the LM Studio team using llama.cpp." }, { "label": "Google Gemma 4 26B A4B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The immutable base config records Gemma4ForConditionalGeneration, bfloat16 source dtype, 30 text layers, 16 attention heads, 8 local KV heads, 2 global KV heads, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, resident vision config, and 262144 max position embeddings." }, { "label": "LM Studio Gemma 4 26B A4B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF/tree/43abd4da26c7b9c1a93c4d96a588c8a5374be8eb", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-26B-A4B-it-QAT-Q4_0.gguf is 14439362752 bytes, exactly matching API gguf.totalFileSize. The sidecar mmproj-gemma-4-26B-A4B-it-QAT-BF16.gguf is 1194827776 bytes and is not the selected main text artifact. A pinned config.json check returned 404, so the profile does not claim repo-local config evidence." }, { "label": "LM Studio Gemma 4 26B A4B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF/resolve/43abd4da26c7b9c1a93c4d96a588c8a5374be8eb/gemma-4-26B-A4B-it-QAT-Q4_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 46 metadata entries and 658 tensors. The selected file is 14.439362752 GB, with tensor payloads starting at byte 15823104. Tensor spans total 14.423538808 GB across 25.233142046B logical elements: Q4_0 13.771929600 GB, Q6_K 0.605552640 GB, and F32 0.046056568 GB. Metadata/tokenizer/header/file overhead accounts for 0.015823944 GB. Non-expert and always-on tensor spans total 1.577172088 GB. Routed expert Q4_0 tensors total 12.846366720 GB across 30 layers and 128 expert indexes, or 0.100362240 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, unified full-attention K/V geometry, and separate sliding-layer K/V projections." }, { "label": "Google Gemma 4 26B A4B IT QAT Q4_0 GGUF audited profile comparison", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf", "source_type": "manual_review", "supports": [ "config_compatible", "weight_format", "kv_adapter" ], "notes": "The LM Studio tensor spans, routed expert bytes, fixed tensor bytes, tensor types, tensor count, metadata count, and GGUF architecture fields match the already audited official Google Q4_0 QAT main GGUF. The only observed selected-main-file difference is 1,312 additional resident overhead bytes in the LM Studio linked file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned LM Studio model card, pinned 404 repo-config check, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, comparison with the existing Google QAT GGUF profile, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for the LM Studio main Q4_0 QAT GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-26b-a4b-it-qat-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit", "title": "LM Studio Gemma 4 26B A4B IT QAT MLX 4-bit", "summary": "Audited memory-side bounds profile for the LM Studio MLX 4-bit QAT package of Gemma 4 26B A4B IT.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model metadata, public Google QAT config comparison, non-QAT Google config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio card metadata lists google/gemma-4-26B-A4B-it-QAT, which is not resolvable through the HF CLI/API in this audit environment. The public google/gemma-4-26B-A4B-it-qat-q4_0-unquantized repo is the matching QAT reference. Manual comparison found no differences in the audited text, vision, context, tied-embedding, expert, and attention geometry fields between this MLX config, the public QAT config, and the non-QAT google/gemma-4-26B-A4B-it config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 15.608614044, "main_resident_weight_gb": 14.467688508, "auxiliary_resident_weight_gb": 1.140925536, "fixed_weight_gb": 1.621321788, "routed_expert_weight_gb": 0.10036224, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges fixed language tensors plus expected distinct routed expert tensors", "auxiliary_scope": "vision_tower and embed_vision tensors are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The Gemma 4 26B config records top_k_experts 8, and the public Gemma card/profile evidence describes 1 shared expert. The MLX headers store dense language_model.model.layers.*.mlp.* and router tensors outside language_model.model.layers.*.experts.*, so those shared/always-on tensors are charged in fixed_weight_gb.", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. Header-derived bytes are authoritative for the bound: routed expert tensors total 12.846366720 GB across 30 layers and 128 stacked expert indexes, or 0.100362240 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma 4 26B profiles. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, not vision prefill or image encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5880250448923184, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-qat-4bit-affine-moe-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, and resident vision tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, and vision prefill are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 4-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the public QAT reference logical total parameter count; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 26B A4B IT QAT MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit f03a4a76828804e3b56758627383fbc7bd32015c, the API records a public/non-gated Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 4-bit, and region:us tags, plus 1191518 downloads. The model card says this is a 4-bit quantized version of gemma-4-26B-A4B-it-QAT using MLX, optimized for Apple Silicon. The API safetensors block reports BF16 1358865998, U32 3222720512, and total 4581586510 storage elements." }, { "label": "LM Studio Gemma 4 26B A4B IT QAT MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit/raw/f03a4a76828804e3b56758627383fbc7bd32015c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, attention_k_eq_v true, 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 128 experts, top_k_experts 8, 16 attention heads, 8 local KV heads, 2 global KV heads, 256 head dimension, 512 global head dimension, 262144 max position embeddings, resident vision config, and an MLX quantization_config with default 4-bit affine group_size 64 plus 120 language MLP/router projection overrides at 8-bit." }, { "label": "Google Gemma 4 26B A4B IT public QAT reference config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of the audited text and vision architecture fields found no differences between this MLX config and the public Google QAT reference config. The public QAT API records BF16 25805936206 and total 26544131376 logical parameters at commit 641f184470aa8554ae7957599a624badc2bf4e57." }, { "label": "Google Gemma 4 26B A4B IT non-QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "attention_pattern", "shared_experts_per_token" ], "notes": "Manual comparison of the same audited architecture fields found no differences between the MLX config and the non-QAT Google instruction-tuned config. The already-audited local profile and public card evidence for this base identify hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "LM Studio Gemma 4 26B A4B IT QAT MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit/raw/f03a4a76828804e3b56758627383fbc7bd32015c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 15608614044 bytes across three shards. Range-read safetensors headers found 1697 tensors totaling the same 15.608614044 GB: BF16 2.717731996 GB and U32 12.890882048 GB. Stored-byte buckets are vision/embed_vision auxiliary tensors 1.140925536 GB, fixed language tensors 1.621321788 GB, and routed expert tensors 12.846366720 GB. Expert tensors are stacked with first dimension 128; all 30 layers have uniform expert bytes of 428212224 bytes, so the exact per-expert-index byte charge is 12.846366720 GB / 128 = 0.100362240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served MLX config, public Google QAT config, non-QAT Google config comparison, model card text, safetensors index, and direct shard header range reads." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights with one routed expert per token. It uses exact MLX package bytes and the correct Gemma 4 26B routing geometry for ordinary text decode." }, { "id": "lmstudio-community--gemma-4-31b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-31B-it-GGUF", "title": "LM Studio Gemma 4 31B IT GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected LM Studio Q4_K_M GGUF artifact of Gemma 4 31B IT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google Gemma 4 31B IT config, file list, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-31B-it. The selected GGUF header records the same Gemma 4 31B dense text geometry as the Google config. The LM Studio GGUF repo does not ship config.json, so the immutable Google config is used for high-level architecture fields and the GGUF header is used for selected artifact details." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 18.687061664, "swept_weight_gb": 18.671230848, "auxiliary_resident_weight_gb": 0.015830816, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31B-it-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, tensor alignment, and file overhead are resident in the selected artifact file but not swept as model tensors; mmproj-gemma-4-31B-it-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_K_M GGUF file selected by HF API gguf.totalFileSize. Header tensor payloads total 18.671229168 GB, with 0.000001680 GB of tensor-alignment padding, while the linked file size is 18.687061664 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The GGUF metadata records four global KV heads and 512 global key/value length. The selected header has k_proj tensors for full-attention layers but no v_proj tensors, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads, 256 key/value length, and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6087517113021983, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_K_M GGUF because HF API gguf.totalFileSize matches gemma-4-31B-it-Q4_K_M.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 31B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-31B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF CLI/API response at commit cc1750d5ecdf85a09a19e2649f343e99cdb19524 records base_model google/gemma-4-31B-it, Apache-2.0 license metadata, region:us, 115645 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, and gguf.totalFileSize 18687061664. The API does not currently expose a pipeline_tag for this repo." }, { "label": "LM Studio Gemma 4 31B GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF/raw/cc1750d5ecdf85a09a19e2649f343e99cdb19524/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "runtime_format" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-31B-it, and says the LM Studio team produced the GGUF quantization with llama.cpp release b8778." }, { "label": "Google Gemma 4 31B IT config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, bfloat16 text config, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, four global KV heads, 512 global key/value length, 16 local KV heads, 256 local key/value length, dense text MLPs, resident Gemma 4 vision config, and 262144 max position embeddings." }, { "label": "LM Studio Gemma 4 31B GGUF file list", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF/tree/cc1750d5ecdf85a09a19e2649f343e99cdb19524", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HF CLI file listing found gemma-4-31B-it-Q4_K_M.gguf at 18687061664 bytes, exactly matching API gguf.totalFileSize. Sibling files are gemma-4-31B-it-Q6_K.gguf at 25201483424 bytes, gemma-4-31B-it-Q8_0.gguf at 32635674272 bytes, and mmproj-gemma-4-31B-it-BF16.gguf at 1200725984 bytes. The mmproj sidecar is not the selected main text artifact." }, { "label": "LM Studio Gemma 4 31B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF/resolve/cc1750d5ecdf85a09a19e2649f343e99cdb19524/gemma-4-31B-it-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 43 metadata entries and 833 tensors. The selected file is 18.687061664 GB, with tensor payloads starting at byte 15830816. Tensor spans total 18.671230848 GB across 30697345596 logical elements: token_embd.weight 1.156055040 GB, blk.* tensors 17.515153280 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Raw tensor payloads split into Q4_K 14.213283840 GB, Q6_K 4.452618240 GB, and F32 0.005327088 GB, with 1680 bytes of tensor-alignment padding. Metadata/tokenizer/header/file overhead accounts for 0.015830816 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, layer KV head array with ten global layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF CLI/API metadata, model card, immutable Google Gemma 4 31B IT config, HF CLI file listing, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-31b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-31B-it-QAT-GGUF", "title": "LM Studio Gemma 4 31B IT QAT GGUF Q4_0", "summary": "Audited memory-side text-decode bounds profile for the LM Studio Q4_0 GGUF artifact of Gemma 4 31B IT QAT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-31B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 31B dense text geometry as the Google QAT unquantized config. The repo ships a main Q4_0 GGUF file plus a separate mmproj sidecar, so the immutable Google QAT unquantized config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 17.651000768, "swept_weight_gb": 17.635168128, "auxiliary_resident_weight_gb": 0.01583264, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31B-it-QAT-Q4_0.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, tensor alignment, and file overhead are resident in the selected artifact file but not swept as model tensors; mmproj-gemma-4-31B-it-QAT-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The profile targets the main Q4_0 GGUF file selected by HF API gguf.totalFileSize. Header tensor payloads total 17.635166448 GB, with 0.000001680 GB of tensor-alignment padding, while the linked file size is 17.651000768 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no output.weight, mmproj, vision, audio, or MTP tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config and GGUF metadata record four global KV heads and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_0 GGUF artifact after any multimodal prefill. The separate mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5750008811934542, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_0 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 31B IT QAT GGUF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-31B-it-QAT-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 5f1655eb80159b7db3f6feb9ce1a9440ab076261 records base_model google/gemma-4-31B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, region:us, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, gguf.totalFileSize 17651000768, and 273894 current downloads." }, { "label": "LM Studio Gemma 4 31B IT QAT GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-QAT-GGUF/raw/5f1655eb80159b7db3f6feb9ce1a9440ab076261/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-31B-it-qat-q4_0-unquantized, and llama.cpp GGUF quantization by the LM Studio team." }, { "label": "Google Gemma 4 31B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, four global KV heads, 512 global head dimension, 16 local KV heads, 256 local head dimension, dense MLPs with enable_moe_block false, resident Gemma 4 vision config, and 262144 max position embeddings." }, { "label": "LM Studio Gemma 4 31B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-QAT-GGUF/tree/5f1655eb80159b7db3f6feb9ce1a9440ab076261", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-31B-it-QAT-Q4_0.gguf is 17651000768 bytes, matching API gguf.totalFileSize. The sidecar mmproj-gemma-4-31B-it-QAT-BF16.gguf is 1200726016 bytes and is not the selected main text artifact." }, { "label": "LM Studio Gemma 4 31B IT QAT Q4_0 GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-QAT-GGUF/resolve/5f1655eb80159b7db3f6feb9ce1a9440ab076261/gemma-4-31B-it-QAT-Q4_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 43 metadata entries and 833 tensors. The selected file is 17.651000768 GB, with tensor payloads starting at byte 15832640. Tensor spans total 17.635168128 GB across 30.697345596B logical elements: token_embd.weight 1.156055040 GB, blk.* tensors 16.479090560 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Tensor payloads split into Q4_0 16.473784320 GB, Q6_K 1.156055040 GB, and F32 0.005327088 GB. Metadata/tokenizer/header/file overhead accounts for 0.015832640 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, a layer KV head array with ten global layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_0 artifact." }, "notes": "Use this profile for the LM Studio main Q4_0 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-e2b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E2B-it-GGUF", "title": "LM Studio Gemma 4 E2B IT GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected LM Studio Q4_K_M GGUF artifact of Gemma 4 E2B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google base config, Transformers PLE documentation, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E2B-it. The selected GGUF header records the same Gemma 4 E2B text geometry as the Google config. The LM Studio GGUF repo does not ship config.json, so the immutable Google config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.647450147, "swept_params_b": 2.298639907, "auxiliary_resident_params_b": 2.34881024, "resident_weight_gb": 3.427877696, "swept_weight_gb": 1.4853019, "auxiliary_resident_weight_gb": 1.942575796, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-E2B-it-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as full matrices per generated token; mmproj-gemma-4-E2B-it-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "Gemma 4 E2B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per_layer_proj_norm.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. The Google config and GGUF header record shared_kv_layers / num_kv_shared_layers 20, so only layers 4, 9, and 14 allocate full-context K/V cache, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records shared_kv_layers 20. Only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7375824565246273, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-ple-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_K_M GGUF because HF API gguf.totalFileSize matches gemma-4-E2B-it-Q4_K_M.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 E2B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 3155b2dec005f7b0a6f721198c26ab11492a38e4 records a public non-gated Apache-2.0 GGUF repo, base_model google/gemma-4-E2B-it, endpoints_compatible, region:us, conversational, 141691 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 4647450147, and gguf.totalFileSize 3427877696." }, { "label": "LM Studio Gemma 4 E2B GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E2B-it, and says LM Studio produced the GGUF quantization using llama.cpp release b8778." }, { "label": "Google Gemma 4 E2B IT config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, one KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, hidden_size_per_layer_input 256, resident vision config, and resident audio config." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "LM Studio Gemma 4 E2B GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-GGUF/tree/3155b2dec005f7b0a6f721198c26ab11492a38e4", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E2B-it-Q4_K_M.gguf is 3427877696 bytes, exactly matching API gguf.totalFileSize. gemma-4-E2B-it-Q6_K.gguf is 3845344064 bytes, gemma-4-E2B-it-Q8_0.gguf is 4967494464 bytes, and mmproj-gemma-4-E2B-it-BF16.gguf is 986833248 bytes. The mmproj sidecar is not the selected main text artifact." }, { "label": "LM Studio Gemma 4 E2B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-GGUF/resolve/3155b2dec005f7b0a6f721198c26ab11492a38e4/gemma-4-E2B-it-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 601 tensors. The selected file is 3.427877696 GB, with tensor payloads starting at byte 15816416. Tensor spans total 3.412060300 GB across 4647450147 logical elements: per_layer_token_embd.weight 1.926758400 GB / 2348810240 logical elements, token_embd.weight 0.330301440 GB, blk.* tensors 1.127467148 GB, per_layer_model_proj/proj_norm tensors 0.027526144 GB, output_norm.weight 0.000006144 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/file overhead accounts for 0.015817396 GB. Swept tensor spans excluding the per-layer token lookup table total 1.485301900 GB across 2298639907 logical elements. Tensor payloads split into Q4_K 0.910835712 GB, Q6_K 2.472529920 GB, BF16 0.027525120 GB, and F32 0.001169548 GB. The header records gemma4.block_count 35, context_length 131072, attention.head_count 8, attention.head_count_kv 1, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, shared_kv_layers 20, embedding_length_per_layer_input 256, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card, immutable Google base config, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-e2b-it-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E2B-it-MLX-4bit", "title": "LM Studio Gemma 4 E2B IT MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Gemma 4 E2B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E2B-it as its base model. Manual comparison found matching text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the audited Google base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 4.339549126, "swept_weight_gb": 2.071972422, "auxiliary_resident_weight_gb": 2.267576704, "resident_parameter_scope": "Google Gemma 4 E2B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard header and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E2B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8470422649272819, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 4-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E2B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 34f17a2e0a341dcbacd9b90ee604521ac88defc7, the API records an Apache-2.0 any-to-any repo with mlx, 4-bit, and region:us tags, base_model google/gemma-4-E2B-it, 242,033 downloads, and safetensors parameters BF16 617,815,619 plus U32 775,979,008, total 1,393,794,627. The card describes this as a 4-bit MLX quantized version of google/gemma-4-E2B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E2B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-4bit/raw/34f17a2e0a341dcbacd9b90ee604521ac88defc7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 4-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit." }, { "label": "Google Gemma 4 E2B IT base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of relevant text, multimodal, context, tied-embedding, dtype, and attention architecture fields found no differences between the MLX config and the audited Google E2B IT config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E2B MLX 4-bit safetensors index and header", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-4bit/raw/34f17a2e0a341dcbacd9b90ee604521ac88defc7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 4,339,549,126 bytes in one safetensors file. Range-reading the safetensors header found a 337,439-byte header and 2,651 tensors totaling 4.339549126 GB: BF16 1.235633094 GB and U32 3.103916032 GB. Header storage elements sum to BF16 617,816,547 and U32 775,979,008, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.226492416 GB, language_model.model.embed_tokens_per_layer 1.321205760 GB, other language_model tensors 1.845480006 GB, audio/embed_audio auxiliary tensors 0.610977280 GB, and vision/embed_vision auxiliary tensors 0.335393664 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 2.071972422 GB. Resident-only multimodal plus per-layer embedding tensors total 2.267576704 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, audited Google base config comparison, model card, safetensors index, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored MLX scales, biases, mixed 8-bit MLP overrides, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e2b-it-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E2B-it-MLX-5bit", "title": "LM Studio Gemma 4 E2B IT MLX 5-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 5-bit quantized Gemma 4 E2B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E2B-it as its base model. Manual comparison found matching checked text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the audited Google base config. The MLX repo adds 5-bit affine quantization metadata, 8-bit language MLP overrides, and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 4.726244294, "swept_weight_gb": 2.311424582, "auxiliary_resident_weight_gb": 2.414819712, "resident_parameter_scope": "Google Gemma 4 E2B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX safetensors header and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX header contains separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E2B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.9225218001114068, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-5bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an MLX quantization_config with default 5-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E2B MLX 5-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-5bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 2f80e230cb30cec567f11f93835102f999423edc, the API records an Apache-2.0 any-to-any repo with mlx, 5-bit, endpoints-compatible, and region:us tags, base_model google/gemma-4-E2B-it, 230120 downloads, and safetensors parameters BF16 617,815,619 plus U32 872,652,800, total 1,490,468,419. The card describes this as a 5-bit MLX quantized version of google/gemma-4-E2B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E2B MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-5bit/raw/2f80e230cb30cec567f11f93835102f999423edc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 5-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit." }, { "label": "Google Gemma 4 E2B IT base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of relevant text, multimodal, context, tied-embedding, dtype, and attention architecture fields found no differences between the MLX config and the audited Google E2B IT config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E2B MLX 5-bit safetensors index and header", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-5bit/raw/2f80e230cb30cec567f11f93835102f999423edc/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 4,726,244,294 bytes in one safetensors file. Range-reading the safetensors header found a 337,440-byte header and 2,651 tensors totaling 4.726244294 GB: BF16 1.235633094 GB and U32 3.490611200 GB. Header storage elements sum to BF16 617,816,547 and U32 872,652,800, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.251658240 GB, language_model.model.embed_tokens_per_layer 1.468006400 GB, other language_model tensors 2.059766342 GB, audio/embed_audio auxiliary tensors 0.611272192 GB, and vision/embed_vision auxiliary tensors 0.335541120 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 2.311424582 GB. Resident-only multimodal plus per-layer embedding tensors total 2.414819712 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, audited Google base config comparison, model card, safetensors index, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense weights and missed stored MLX scales, biases, mixed 8-bit MLP overrides, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e2b-it-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E2B-it-MLX-6bit", "title": "LM Studio Gemma 4 E2B IT MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Gemma 4 E2B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E2B-it as its base model. Manual comparison found matching checked text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the audited Google base config. The MLX repo adds 6-bit affine quantization metadata, 8-bit language MLP overrides, and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 5.112939462, "swept_weight_gb": 2.404076102, "auxiliary_resident_weight_gb": 2.70886336, "resident_parameter_scope": "Google Gemma 4 E2B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX safetensors header and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX header contains separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E2B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.9980013352955318, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an MLX quantization_config with default 6-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E2B MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 182790d74c2967b8bba0a458d507cfcb0482e8a6, the API records an Apache-2.0 any-to-any repo with mlx, 6-bit, endpoints-compatible, and region:us tags, base_model google/gemma-4-E2B-it, 230149 downloads, and safetensors parameters BF16 617,815,619 plus U32 969,326,592, total 1,587,142,211. The card describes this as a 6-bit MLX quantized version of google/gemma-4-E2B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E2B MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-6bit/raw/182790d74c2967b8bba0a458d507cfcb0482e8a6/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 6-bit affine group_size 64 plus 105 language MLP projection overrides at 8-bit." }, { "label": "Google Gemma 4 E2B IT base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of relevant text, multimodal, context, tied-embedding, dtype, and attention architecture fields found no differences between the MLX config and the audited Google E2B IT config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E2B MLX 6-bit safetensors index and header", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-6bit/raw/182790d74c2967b8bba0a458d507cfcb0482e8a6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 5,112,939,462 bytes in one safetensors file. Range-reading the safetensors header found a 337,440-byte header and 2,651 tensors totaling 5.112939462 GB: BF16 1.235633094 GB and U32 3.877306368 GB. Header storage elements sum to BF16 617,816,547 and U32 969,326,592, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.301989888 GB, language_model.model.embed_tokens_per_layer 1.761607680 GB, other language_model tensors 2.102086214 GB, audio/embed_audio auxiliary tensors 0.611567104 GB, and vision/embed_vision auxiliary tensors 0.335688576 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 2.404076102 GB. Resident-only multimodal plus per-layer embedding tensors total 2.708863360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, audited Google base config comparison, model card, safetensors index, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense weights and missed stored MLX scales, biases, mixed 8-bit MLP overrides, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e2b-it-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E2B-it-MLX-8bit", "title": "LM Studio Gemma 4 E2B IT MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Gemma 4 E2B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E2B-it as its base model. Manual comparison found matching checked text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the audited Google base config. The MLX repo adds 8-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.123178979, "swept_params_b": 2.298639651, "auxiliary_resident_params_b": 2.824539328, "resident_weight_gb": 5.886329798, "swept_weight_gb": 2.589379142, "auxiliary_resident_weight_gb": 3.296950656, "resident_parameter_scope": "Google Gemma 4 E2B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard headers and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. With num_kv_shared_layers 20, only layers 4, 9, and 14 are non-shared K/V producers, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "With num_kv_shared_layers 20, only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E2B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.1489604056637819, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an MLX quantization_config with default 8-bit affine group_size 64. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E2B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit df2b6aabee32f10605d5333775c18ff10b9d60e5, the API records an Apache-2.0 any-to-any repo with mlx, 8-bit, and region:us tags, base_model google/gemma-4-E2B-it, 231439 downloads, and safetensors parameters BF16 617,815,619 plus U32 1,162,674,176, total 1,780,489,795. The card describes this as an 8-bit MLX quantized version of google/gemma-4-E2B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E2B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-8bit/raw/df2b6aabee32f10605d5333775c18ff10b9d60e5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, 1 KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 8-bit affine group_size 64." }, { "label": "Google Gemma 4 E2B IT base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of relevant text, multimodal, context, tied-embedding, dtype, and attention architecture fields found no differences between the MLX config and the audited Google E2B IT config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E2B MLX 8-bit safetensors index and headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-8bit/raw/df2b6aabee32f10605d5333775c18ff10b9d60e5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 5,886,329,798 bytes across two safetensors files. Range-reading both safetensors headers found 335,818 header bytes and 2,651 tensors totaling 5.886329798 GB: BF16 1.235633094 GB and U32 4.650696704 GB. Header storage elements sum to BF16 617,816,547 and U32 1,162,674,176, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.402653184 GB, language_model.model.embed_tokens_per_layer 2.348810240 GB, other language_model tensors 2.186725958 GB, audio/embed_audio auxiliary tensors 0.612156928 GB, and vision/embed_vision auxiliary tensors 0.335983488 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 2.589379142 GB. Resident-only multimodal plus per-layer embedding tensors total 3.296950656 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served MLX config, audited Google base config comparison, model card, safetensors index, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit dense weights and missed stored MLX scales, biases, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E4B-it-GGUF", "title": "LM Studio Gemma 4 E4B IT GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected LM Studio Q4_K_M GGUF artifact of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google base config, Transformers Gemma 4 PLE documentation, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E4B-it. The selected GGUF header records the same Gemma 4 E4B text geometry as the Google config. The LM Studio GGUF repo does not ship config.json, so the immutable Google config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.51806929, "swept_params_b": 4.699497002, "auxiliary_resident_params_b": 2.818572288, "resident_weight_gb": 5.335289664, "swept_weight_gb": 3.007356224, "auxiliary_resident_weight_gb": 2.32793344, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-4-E4B-it-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges GGUF tensor spans in the selected main artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; mmproj-gemma-4-E4B-it-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "Gemma 4 E4B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per-layer projection tensors inside blk.*, normal language blocks, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The Google config records global_head_dim 512 and attention_k_eq_v false, and the selected GGUF header contains separate attn_k and attn_v tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7096622095644453, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_K_M GGUF because HF API gguf.totalFileSize matches gemma-4-E4B-it-Q4_K_M.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "LM Studio Gemma 4 E4B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 53a691ddc52708042c56f80cdaf47f8a1daf051e records base_model google/gemma-4-E4B-it, Apache-2.0 license, region:us, 770548 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 7518069290, and gguf.totalFileSize 5335289664." }, { "label": "LM Studio Gemma 4 E4B GGUF model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E4B-it, and says LM Studio produced the GGUF quantization using llama.cpp release b8778." }, { "label": "Google Gemma 4 E4B IT config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "LM Studio Gemma 4 E4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-GGUF/tree/53a691ddc52708042c56f80cdaf47f8a1daf051e", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E4B-it-Q4_K_M.gguf is 5335289664 bytes, exactly matching API gguf.totalFileSize. Other siblings are Q6_K 6217260864 bytes, Q8_0 8031240000 bytes, and mmproj-gemma-4-E4B-it-BF16.gguf 991551840 bytes; those are not the selected main artifact." }, { "label": "LM Studio Gemma 4 E4B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-GGUF/resolve/53a691ddc52708042c56f80cdaf47f8a1daf051e/gemma-4-E4B-it-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 16MB range-read of the GGUF v3 header found 42 metadata entries and 720 tensors. The selected file is 5.335289664 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15823360. Tensor spans total 5.319466304 GB across 7518069290 logical elements. per_layer_token_embd.weight is 2.312110080 GB / 2818572288 logical elements and is resident-only for this ordinary text-decode profile. Swept tensor spans excluding that lookup table total 3.007356224 GB / 4699497002 logical elements. token_embd.weight is 0.550502400 GB and no output.weight tensor is stored, so token_embd.weight remains swept as tied output-projection traffic. The header records gemma4.block_count 42, context_length 131072, attention.head_count 8, attention.head_count_kv 2, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, embedding_length_per_layer_input 256, and no mmproj, vision, or audio tensors in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google config, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--gemma-4-e4b-it-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E4B-it-MLX-4bit", "title": "LM Studio Gemma 4 E4B IT MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Gemma 4 E4B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E4B-it as its base model. Manual comparison found matching text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the Google base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 6.828932052, "swept_weight_gb": 4.295787092, "auxiliary_resident_weight_gb": 2.53314496, "resident_parameter_scope": "Google Gemma 4 E4B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard headers and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E4B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8540267149603031, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 4-bit affine group_size 64 plus 126 language MLP projection overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E4B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit fa6f15978ba53de9ff3a95ce2821deb0ca4a15f0, the API records an Apache-2.0 any-to-any repo with mlx and 4-bit tags, base_model google/gemma-4-E4B-it, 1,681,799 downloads, and safetensors parameters BF16 707,861,066 plus U32 1,353,302,016, total 2,061,163,082. The card describes this as a 4-bit MLX quantized version of google/gemma-4-E4B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E4B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-4bit/raw/fa6f15978ba53de9ff3a95ce2821deb0ca4a15f0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 4-bit affine group_size 64 plus 126 language MLP projection overrides at 8-bit." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of 33 relevant architecture fields found no differences between the MLX config and the Google base config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E4B MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-4bit/raw/fa6f15978ba53de9ff3a95ce2821deb0ca4a15f0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 6828932052 bytes across two shards. Range-read safetensors headers found 2896 tensors totaling 6.828932052 GB: BF16 1.415723988 GB and U32 5.413208064 GB. Header storage elements sum to BF16 707,861,994 and U32 1,353,302,016, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.37748736 GB, language_model.model.embed_tokens_per_layer 1.585446912 GB, other language_model tensors 3.918299732 GB, audio/embed_audio auxiliary tensors 0.611862016 GB, and vision/embed_vision auxiliary tensors 0.335836032 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 4.295787092 GB. Resident-only multimodal plus per-layer embedding tensors total 2.53314496 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored MLX scales, biases, mixed 8-bit MLP overrides, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e4b-it-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E4B-it-MLX-5bit", "title": "LM Studio Gemma 4 E4B IT MLX 5-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 5-bit quantized Gemma 4 E4B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E4B-it as its base model. Manual comparison found matching text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the Google base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 7.356480468, "swept_weight_gb": 4.470276692, "auxiliary_resident_weight_gb": 2.886203776, "resident_parameter_scope": "Google Gemma 4 E4B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard headers and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E4B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.9200019563696888, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-5bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 5-bit affine group_size 64 plus 126 language MLP projection entries set to 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E4B MLX 5-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-5bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit cdb020657e3404e15a830bb61c6128dd77a2f6f3, the API records an Apache-2.0 any-to-any repo with mlx and 5-bit tags, base_model google/gemma-4-E4B-it, 1,623,424 downloads, and safetensors parameters BF16 707,861,066 plus U32 1,485,189,120, total 2,193,050,186. The card describes this as a 5-bit MLX quantized version of google/gemma-4-E4B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E4B MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-5bit/raw/cdb020657e3404e15a830bb61c6128dd77a2f6f3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 5-bit affine group_size 64 plus 126 language MLP projection entries set to 8-bit." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of 33 relevant architecture fields found no differences between the MLX config and the Google base config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E4B MLX 5-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-5bit/raw/cdb020657e3404e15a830bb61c6128dd77a2f6f3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 7356480468 bytes across two shards. Range-read safetensors headers found 2896 tensors totaling 7.356480468 GB: BF16 1.415723988 GB and U32 5.94075648 GB. Header storage elements sum to BF16 707,861,994 and U32 1,485,189,120, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.46137344 GB, language_model.model.embed_tokens_per_layer 1.937768448 GB, other language_model tensors 4.008903252 GB, audio/embed_audio auxiliary tensors 0.612353536 GB, and vision/embed_vision auxiliary tensors 0.336081792 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 4.470276692 GB. Resident-only multimodal plus per-layer embedding tensors total 2.886203776 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 and overcounted stored MLX weight traffic for ordinary text decode. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e4b-it-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E4B-it-MLX-6bit", "title": "LM Studio Gemma 4 E4B IT MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Gemma 4 E4B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E4B-it as its base model. Manual comparison found matching text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the Google base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 7.884028884, "swept_weight_gb": 4.644766292, "auxiliary_resident_weight_gb": 3.239262592, "resident_parameter_scope": "Google Gemma 4 E4B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard headers and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E4B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.9859771977790746, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 6-bit affine group_size 64 plus 126 language MLP projection entries set to 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E4B MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit ff79ace56197c89ba21729e53c6378a96fe6367c, the API records an Apache-2.0 any-to-any repo with mlx and 6-bit tags, base_model google/gemma-4-E4B-it, 1,622,196 downloads, and safetensors parameters BF16 707,861,066 plus U32 1,617,076,224, total 2,324,937,290. The card describes this as a 6-bit MLX quantized version of google/gemma-4-E4B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E4B MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-6bit/raw/ff79ace56197c89ba21729e53c6378a96fe6367c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 6-bit affine group_size 64 plus 126 language MLP projection entries set to 8-bit." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of 32 relevant architecture fields found no differences between the MLX config and the Google base config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E4B MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-6bit/raw/ff79ace56197c89ba21729e53c6378a96fe6367c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 7884028884 bytes across two shards. Range-read safetensors headers found 2896 tensors totaling 7.884028884 GB: BF16 1.415723988 GB and U32 6.468304896 GB. Header storage elements sum to BF16 707,861,994 and U32 1,617,076,224, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.54525952 GB, language_model.model.embed_tokens_per_layer 2.290089984 GB, other language_model tensors 4.099506772 GB, audio/embed_audio auxiliary tensors 0.612845056 GB, and vision/embed_vision auxiliary tensors 0.336327552 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 4.644766292 GB. Resident-only multimodal plus per-layer embedding tensors total 3.239262592 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 and overcounted stored MLX weight traffic for ordinary text decode. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--gemma-4-e4b-it-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/gemma-4-E4B-it-MLX-8bit", "title": "LM Studio Gemma 4 E4B IT MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Gemma 4 E4B IT artifact.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records google/gemma-4-E4B-it as its base model. Manual comparison found matching text, vision, audio, context, tied-embedding, and attention geometry fields between the MLX config and the Google base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the model architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.996157418, "swept_params_b": 4.699496746, "auxiliary_resident_params_b": 3.296660672, "resident_weight_gb": 8.939125716, "swept_weight_gb": 4.993745492, "auxiliary_resident_weight_gb": 3.945380224, "resident_parameter_scope": "Google Gemma 4 E4B IT logical parameter split with direct MLX safetensors stored-byte totals", "swept_parameter_scope": "language_model safetensors headers excluding embed_tokens_per_layer but including the tied standard embed_tokens output projection", "auxiliary_scope": "audio_tower, embed_audio, vision_tower, embed_vision, and language_model.model.embed_tokens_per_layer tensors are resident for the multimodal PLE package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements rather than base logical model parameters. The parameter fields follow the audited Google BF16 logical split, while resident_weight_gb and swept_weight_gb are exact stored-byte sums from the MLX shard headers and are authoritative for the bound. The index has no lm_head tensor, so language_model.model.embed_tokens is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and the MLX headers contain separate k_proj and v_proj tensors." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The MLX quantization metadata applies to stored weights; no KV-cache quantization scheme is declared, so this profile charges BF16 KV cache bytes like the base Gemma E4B profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.1179276805978458, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized auxiliary tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 model dtype and an MLX quantization_config with default 8-bit affine group_size 64 plus 126 language MLP projection entries also set to 8-bit. weight_bytes_per_param records resident stored bytes divided by the audited base logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Gemma 4 E4B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit c63b2f9519d800e591cb331e5c19c021f66bf79a, the API records an Apache-2.0 any-to-any repo with mlx and 8-bit tags, base_model google/gemma-4-E4B-it, 1,632,405 downloads, and safetensors parameters BF16 707,861,066 plus U32 1,880,850,432, total 2,588,711,498. The card describes this as an 8-bit MLX quantized version of google/gemma-4-E4B-it optimized for Apple Silicon." }, { "label": "LM Studio Gemma 4 E4B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-8bit/raw/c63b2f9519d800e591cb331e5c19c021f66bf79a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, resident audio config, and an MLX quantization_config with default 8-bit affine group_size 64 plus 126 language MLP projection entries also set to 8-bit." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of 33 relevant architecture fields found no differences between the MLX config and the Google base config after excluding quantization metadata and repository bookkeeping." }, { "label": "LM Studio Gemma 4 E4B MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-8bit/raw/c63b2f9519d800e591cb331e5c19c021f66bf79a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The index records total_size 8939125716 bytes across two shards. Range-read safetensors headers found 2896 tensors totaling 8.939125716 GB: BF16 1.415723988 GB and U32 7.523401728 GB. Header storage elements sum to BF16 707,861,994 and U32 1,880,850,432, 928 BF16 scalar elements above the API count. Stored-byte buckets are language_model.model.embed_tokens 0.71303168 GB, language_model.model.embed_tokens_per_layer 2.994733056 GB, other language_model tensors 4.280713812 GB, audio/embed_audio auxiliary tensors 0.613828096 GB, and vision/embed_vision auxiliary tensors 0.336819072 GB. The index has no lm_head tensors. Ordinary text swept traffic is language_model excluding embed_tokens_per_layer but including the tied standard embedding/output projection, totaling 4.993745492 GB. Resident-only multimodal plus per-layer embedding tensors total 3.945380224 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit dense weights and undercounted stored MLX scales, biases, multimodal tensors, and per-layer embedding storage. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--glm-4-6v-flash-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/GLM-4.6V-Flash-MLX-4bit", "title": "LM Studio GLM 4.6V Flash MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized GLM-4.6V-Flash multimodal artifact.", "model_family": "glm4v-dense-vlm-mlx", "base_model_proof": { "base_model": "zai-org/GLM-4.6V-Flash", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, current base config, direct MLX safetensors header grouping, and mlx-vlm runtime review", "config_compatible": true, "notes": "The LM Studio repo records zai-org/GLM-4.6V-Flash as its base model. The pinned target config matches the current base config across the checked memory-relevant text and vision geometry: Glm4vForConditionalGeneration, glm4v_text, 40 text layers, hidden size 4096, intermediate size 13696, 32 attention heads, 2 KV heads, 128 head dimension, 131072 max positions, untied embeddings, and the GLM4V vision tower metadata. The target repo adds MLX 4-bit affine quantization metadata and packed U32/BF16 storage tensors." }, "architecture": { "canonical_architecture_id": "glm-4-6v-flash-mlx", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.292777472, "swept_params_b": 8.779522048, "auxiliary_resident_params_b": 1.513255424, "resident_weight_gb": 7.073866752, "swept_weight_gb": 4.93969408, "auxiliary_resident_weight_gb": 2.134172672, "resident_parameter_scope": "logical GLM4V package parameter count from the base config and stale index total, with exact resident bytes taken from the actual two downloadable MLX safetensors shard headers", "swept_parameter_scope": "ordinary text decode includes language_model.model.layers.0-39, language_model.model.norm, and language_model.lm_head tensors from the sanitized mlx-vlm weight names", "auxiliary_scope": "language_model.model.embed_tokens tensors and all vision_tower tensors are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. The checked-in model.safetensors.index.json is stale for this revision because it names four shards that return 404; the actual downloadable files are model-00001-of-00002.safetensors and model-00002-of-00002.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The target and base text configs record 40 layers, 2 KV heads, 128 head dimension, use_cache true, and no sliding-window or recurrent-state text cache. The mlx-vlm GLM4V attention implementation calls cache.update_and_fetch(keys, values), and the default KVCache stores both keys and values." }, "notes": "Glm4vForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill; it does not model vision encoder or projector throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.6872651013046209, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-glm4v-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head side tensors, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 4-bit affine quantization with group_size 64. weight_bytes_per_param is the exact resident payload divided by the logical package parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio GLM 4.6V Flash MLX 4-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit f3626bba332c159be847ee1a4e68a5fcbde915ae, the live API records a public non-gated MIT image-text-to-text repo with transformers, safetensors, glm4v, mlx, conversational, zh/en, endpoints_compatible, 4-bit, region:us, and base_model zai-org/GLM-4.6V-Flash tags. Current downloads are 111225." }, { "label": "LM Studio GLM 4.6V Flash MLX 4-bit model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/raw/f3626bba332c159be847ee1a4e68a5fcbde915ae/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the Z.ai GLM-4.6V-Flash base and says the package is a 4-bit MLX quantization provided by the LM Studio team using mlx_vlm, optimized for Apple Silicon." }, { "label": "LM Studio GLM 4.6V Flash MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/raw/f3626bba332c159be847ee1a4e68a5fcbde915ae/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Glm4vForConditionalGeneration, glm4v, tie_word_embeddings false, 4-bit affine MLX quantization with group_size 64, bfloat16 text dtype, 40 text layers, hidden_size 4096, intermediate_size 13696, 32 attention heads, 2 KV heads, 128 effective head dimension, 131072 max position embeddings, use_cache true, vocab_size 151552, and GLM4V vision metadata." }, { "label": "GLM 4.6V Flash BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "The current base config records the same checked memory-relevant GLM4V architecture fields as the LM Studio target config and exposes BF16 safetensors total 10292777472 parameters in the live API metadata. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio GLM 4.6V Flash MLX 4-bit safetensors headers", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/tree/f3626bba332c159be847ee1a4e68a5fcbde915ae", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Direct HEAD checks and safetensors header range reads of the actual linked files found model-00001-of-00002.safetensors at 5.367706688 GB and model-00002-of-00002.safetensors at 1.706311080 GB. Their tensor payload totals 7.073866752 GB across 1188 tensors, with 0.000151016 GB safetensors header/container overhead. Payload bytes split into U32 4.699717632 GB and BF16 2.374149120 GB. Ordinary swept text tensors, language_model.model.layers.0-39 plus language_model.model.norm and language_model.lm_head, sum to 4.939694080 GB. language_model.model.embed_tokens is 0.349175808 GB resident-only, vision_tower is 1.784996864 GB resident-only, and the resident-only total is 2.134172672 GB." }, { "label": "LM Studio GLM 4.6V Flash MLX 4-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/raw/f3626bba332c159be847ee1a4e68a5fcbde915ae/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The checked-in index metadata records total_size 10292777472 and maps 704 weights to model-00001-of-00004.safetensors through model-00004-of-00004.safetensors, but direct HEAD requests for the four indexed shard names return 404. The Hub file listing exposes the two actual shards used in this profile, so the index is treated as stale naming metadata, not as the selected artifact byte source." }, { "label": "mlx-vlm GLM4V language and cache implementation", "url": "https://github.com/Blaizzy/mlx-vlm/blob/0dd968334619594325128e4325ba2d8fe1885064/mlx_vlm/models/glm4v/language.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GLM4VModel instantiates range(config.num_hidden_layers), so ordinary decode uses 40 text layers. Glm4vAttention projects Q, K, and V separately, reshapes K/V to num_key_value_heads, applies multimodal RoPE, then calls cache.update_and_fetch(keys, values). The pinned mlx-vlm default KVCache stores both K and V arrays, so Bounds Engine v1 models full-context BF16 K/V cache traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, model card, pinned served MLX config, current BF16 base config comparison, stale index review, direct two-shard safetensors header byte grouping, and pinned mlx-vlm GLM4V/cache source review." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps resident visual/embedding tensors separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--glm-4-6v-flash-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/GLM-4.6V-Flash-MLX-6bit", "title": "LM Studio GLM 4.6V Flash MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized GLM-4.6V-Flash multimodal artifact.", "model_family": "glm4v-dense-vlm-mlx", "base_model_proof": { "base_model": "zai-org/GLM-4.6V-Flash", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, current base config, direct MLX safetensors header grouping, and mlx-vlm runtime review", "config_compatible": true, "notes": "The LM Studio repo records zai-org/GLM-4.6V-Flash as its base model. The pinned target config matches the current base config across the checked memory-relevant text and vision geometry: Glm4vForConditionalGeneration, glm4v_text, 40 text layers, hidden size 4096, intermediate size 13696, 32 attention heads, 2 KV heads, 128 head dimension, 131072 max positions, untied embeddings, and the GLM4V vision tower metadata. The target repo adds MLX 6-bit affine quantization metadata and packed U32/BF16 storage tensors." }, "architecture": { "canonical_architecture_id": "glm-4-6v-flash-mlx", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.292777472, "swept_params_b": 8.779522048, "auxiliary_resident_params_b": 1.513255424, "resident_weight_gb": 9.423725568, "swept_weight_gb": 7.134363648, "auxiliary_resident_weight_gb": 2.28936192, "resident_parameter_scope": "logical GLM4V package parameter count from the base config and stale index total, with exact resident bytes taken from the actual two downloadable MLX safetensors shard headers", "swept_parameter_scope": "ordinary text decode includes language_model.model.layers.0-39, language_model.model.norm, and all language_model.lm_head tensors from the sanitized mlx-vlm weight names", "auxiliary_scope": "language_model.model.embed_tokens tensors and all vision_tower tensors are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. The checked-in model.safetensors.index.json is stale for this revision because it names four shards that return 404; the actual downloadable files are model-00001-of-00002.safetensors and model-00002-of-00002.safetensors. Unlike the 4-bit and 8-bit siblings, this 6-bit package stores lm_head.weight plus lm_head.scales and lm_head.biases, so all three lm_head tensors are included in ordinary swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The target and base text configs record 40 layers, 2 KV heads, 128 head dimension, use_cache true, and no sliding-window or recurrent-state text cache. The mlx-vlm GLM4V attention implementation calls cache.update_and_fetch(keys, values), and the default KVCache stores both keys and values." }, "notes": "Glm4vForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill; it does not model vision encoder or projector throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.9155668228168607, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-glm4v-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head side tensors, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 6-bit affine quantization with group_size 64. weight_bytes_per_param is the exact resident payload divided by the logical package parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio GLM 4.6V Flash MLX 6-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 485d12f4f975842648828f8904939650f130a8d5, the live API records a public non-gated MIT image-text-to-text repo with transformers, safetensors, glm4v, mlx, conversational, zh/en, endpoints_compatible, 6-bit, region:us, and base_model zai-org/GLM-4.6V-Flash tags. Current downloads are 108813." }, { "label": "LM Studio GLM 4.6V Flash MLX 6-bit model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/raw/485d12f4f975842648828f8904939650f130a8d5/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the Z.ai GLM-4.6V-Flash base and says the package is a 6-bit MLX quantization provided by the LM Studio team using mlx_vlm, optimized for Apple Silicon." }, { "label": "LM Studio GLM 4.6V Flash MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/raw/485d12f4f975842648828f8904939650f130a8d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Glm4vForConditionalGeneration, glm4v, tie_word_embeddings false, 6-bit affine MLX quantization with group_size 64, bfloat16 text dtype, 40 text layers, hidden_size 4096, intermediate_size 13696, 32 attention heads, 2 KV heads, 128 effective head dimension, 131072 max position embeddings, use_cache true, vocab_size 151552, and GLM4V vision metadata." }, { "label": "GLM 4.6V Flash BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "The current base config records the same checked memory-relevant GLM4V architecture fields as the LM Studio target config and exposes BF16 safetensors total 10292777472 parameters in the live API metadata. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio GLM 4.6V Flash MLX 6-bit safetensors headers", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/tree/485d12f4f975842648828f8904939650f130a8d5", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Direct HEAD checks and safetensors header range reads of the actual linked files found model-00001-of-00002.safetensors at 5.301915092 GB and model-00002-of-00002.safetensors at 4.121961790 GB. Their tensor payload totals 9.423725568 GB across 1188 tensors, with 0.000151314 GB safetensors header/container overhead. Payload bytes split into U32 7.049576448 GB and BF16 2.374149120 GB. Ordinary swept text tensors, language_model.model.layers.0-39 plus language_model.model.norm plus language_model.lm_head.weight/scales/biases, sum to 7.134363648 GB. language_model.model.embed_tokens is 0.504365056 GB resident-only, vision_tower is 1.784996864 GB resident-only, and the resident-only total is 2.289361920 GB." }, { "label": "LM Studio GLM 4.6V Flash MLX 6-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/raw/485d12f4f975842648828f8904939650f130a8d5/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The checked-in index metadata records total_size 10292777472 and maps 704 weights to model-00001-of-00004.safetensors through model-00004-of-00004.safetensors, but direct HEAD requests for the four indexed shard names return 404. The Hub file listing exposes the two actual shards used in this profile, so the index is treated as stale naming metadata, not as the selected artifact byte source." }, { "label": "mlx-vlm GLM4V language and cache implementation", "url": "https://github.com/Blaizzy/mlx-vlm/blob/0dd968334619594325128e4325ba2d8fe1885064/mlx_vlm/models/glm4v/language.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GLM4VModel instantiates range(config.num_hidden_layers), so ordinary decode uses 40 text layers. Glm4vAttention projects Q, K, and V separately, reshapes K/V to num_key_value_heads, applies multimodal RoPE, then calls cache.update_and_fetch(keys, values). The pinned mlx-vlm default KVCache stores both K and V arrays, so Bounds Engine v1 models full-context BF16 K/V cache traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, model card, pinned served MLX config, current BF16 base config comparison, stale index review, direct two-shard safetensors header byte grouping, and pinned mlx-vlm GLM4V/cache source review." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps resident visual/embedding tensors separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--glm-4-6v-flash-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/GLM-4.6V-Flash-MLX-8bit", "title": "LM Studio GLM 4.6V Flash MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized GLM-4.6V-Flash multimodal artifact.", "model_family": "glm4v-dense-vlm-mlx", "base_model_proof": { "base_model": "zai-org/GLM-4.6V-Flash", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, current base config, direct MLX safetensors header grouping, and mlx-vlm runtime review", "config_compatible": true, "notes": "The LM Studio repo records zai-org/GLM-4.6V-Flash as its base model. The pinned target config matches the current base config across the checked memory-relevant text and vision geometry: Glm4vForConditionalGeneration, glm4v_text, 40 text layers, hidden size 4096, intermediate size 13696, 32 attention heads, 2 KV heads, 128 head dimension, 131072 max positions, untied embeddings, and the GLM4V vision tower metadata. The target repo adds MLX 8-bit affine quantization metadata and packed U32/BF16 storage tensors." }, "architecture": { "canonical_architecture_id": "glm-4-6v-flash-mlx", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.292777472, "swept_params_b": 8.779522048, "auxiliary_resident_params_b": 1.513255424, "resident_weight_gb": 11.773584384, "swept_weight_gb": 9.329033216, "auxiliary_resident_weight_gb": 2.444551168, "resident_parameter_scope": "logical GLM4V package parameter count from the base config and stale index total, with exact resident bytes taken from the actual three downloadable MLX safetensors shard headers", "swept_parameter_scope": "ordinary text decode includes language_model.model.layers.0-39, language_model.model.norm, and language_model.lm_head tensors from the sanitized mlx-vlm weight names", "auxiliary_scope": "language_model.model.embed_tokens tensors and all vision_tower tensors are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. The checked-in model.safetensors.index.json is stale for this revision because it names four shards that return 404; the actual downloadable files are model-00001-of-00003.safetensors, model-00002-of-00003.safetensors, and model-00003-of-00003.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The target and base text configs record 40 layers, 2 KV heads, 128 head dimension, use_cache true, and no sliding-window or recurrent-state text cache. The mlx-vlm GLM4V attention implementation calls cache.update_and_fetch(keys, values), and the default KVCache stores both keys and values." }, "notes": "Glm4vForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill; it does not model vision encoder or projector throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.1438685443291006, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-glm4v-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head side tensors, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 8-bit affine quantization with group_size 64. weight_bytes_per_param is the exact resident payload divided by the logical package parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio GLM 4.6V Flash MLX 8-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d, the live API records a public non-gated MIT image-text-to-text repo with transformers, safetensors, glm4v, mlx, conversational, zh/en, endpoints_compatible, 8-bit, region:us, and base_model zai-org/GLM-4.6V-Flash tags. Current downloads are 109777." }, { "label": "LM Studio GLM 4.6V Flash MLX 8-bit model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/raw/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the Z.ai GLM-4.6V-Flash base and says the package is an 8-bit MLX quantization provided by the LM Studio team using mlx_vlm, optimized for Apple Silicon." }, { "label": "LM Studio GLM 4.6V Flash MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/raw/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Glm4vForConditionalGeneration, glm4v, tie_word_embeddings false, 8-bit affine MLX quantization with group_size 64, bfloat16 text dtype, 40 text layers, hidden_size 4096, intermediate_size 13696, 32 attention heads, 2 KV heads, 128 effective head dimension, 131072 max position embeddings, use_cache true, vocab_size 151552, and GLM4V vision metadata." }, { "label": "GLM 4.6V Flash BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "The current base config records the same checked memory-relevant GLM4V architecture fields as the LM Studio target config and exposes BF16 safetensors total 10292777472 parameters in the live API metadata. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio GLM 4.6V Flash MLX 8-bit safetensors headers", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/tree/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Direct HEAD checks and safetensors header range reads of the actual linked files found model-00001-of-00003.safetensors at 5.300083366 GB, model-00002-of-00003.safetensors at 5.320970853 GB, and model-00003-of-00003.safetensors at 1.152681589 GB. Their tensor payload totals 11.773584384 GB across 1188 tensors, with 0.000151424 GB safetensors header/container overhead. Payload bytes split into U32 9.399435264 GB and BF16 2.374149120 GB. Ordinary swept text tensors, language_model.model.layers.0-39 plus language_model.model.norm and language_model.lm_head, sum to 9.329033216 GB. language_model.model.embed_tokens is 0.659554304 GB resident-only, vision_tower is 1.784996864 GB resident-only, and the resident-only total is 2.444551168 GB." }, { "label": "LM Studio GLM 4.6V Flash MLX 8-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/raw/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The checked-in index metadata records total_size 10292777472 and maps 704 weights to model-00001-of-00004.safetensors through model-00004-of-00004.safetensors, but direct HEAD requests for the four indexed shard names return 404. The Hub file listing exposes the three actual shards used in this profile, so the index is treated as stale naming metadata, not as the selected artifact byte source." }, { "label": "mlx-vlm GLM4V language and cache implementation", "url": "https://github.com/Blaizzy/mlx-vlm/blob/0dd968334619594325128e4325ba2d8fe1885064/mlx_vlm/models/glm4v/language.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GLM4VModel instantiates range(config.num_hidden_layers), so ordinary decode uses 40 text layers. Glm4vAttention projects Q, K, and V separately, reshapes K/V to num_key_value_heads, applies multimodal RoPE, then calls cache.update_and_fetch(keys, values). The pinned mlx-vlm default KVCache stores both K and V arrays, so Bounds Engine v1 models full-context BF16 K/V cache traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, model card, pinned served MLX config, current BF16 base config comparison, stale index review, direct three-shard safetensors header byte grouping, and pinned mlx-vlm GLM4V/cache source review." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps resident visual/embedding tensors separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--glm-4-7-flash-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/GLM-4.7-Flash-MLX-6bit", "title": "LM Studio GLM 4.7 Flash MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized GLM-4.7-Flash artifact.", "model_family": "glm4-moe-lite", "base_model_proof": { "base_model": "zai-org/GLM-4.7-Flash", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records zai-org/GLM-4.7-Flash as its base model. Manual comparison found no differences across 34 checked architecture fields between the MLX config, the audited 8-bit MLX sibling config, and the audited BF16 source config. The MLX repo adds 6-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the GLM4 MoE Lite architecture." }, "architecture": { "canonical_architecture_id": "glm-4-7-flash", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.336479232, "main_resident_weight_gb": 24.058934272, "auxiliary_resident_weight_gb": 0.27754496, "fixed_weight_gb": 1.485190144, "routed_expert_weight_gb": 0.352714752, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16_f32", "traffic_scope": "ordinary MLX text decode through model.layers.0-46, model.norm, and lm_head, excluding model.embed_tokens resident-only lookup tensors", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each ordinary generated text token; the MLX artifact has no model.layers.47 tensors despite the config declaring one next-token-prediction layer", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package stores MLX U32 packed tensors, BF16 quantization metadata, and tiny F32 router correction biases. Routed experts are packed under mlp.switch_mlp with a leading 64-expert dimension; switch_mlp tensors sum to 22.573744128 GB, or 0.352714752 GB per expert group across layers 1-46." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.054144, "notes": "MLX GLM4 MoE Lite latent-cache coefficient: 47 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.054144, "notes": "Decode reads the same cached kv_latent plus RoPE key state per active context token in this v1 memory-traffic approximation." }, "notes": "The audited MLX implementation caches kv_latent with rank 512 and k_pe with RoPE dimension 64 through cache.update_and_fetch, rather than storing expanded full K/V heads. This profile is runtime-specific to the MLX artifact; the BF16 Transformers profile remains expanded-K/V." }, "notes": "The served config records 47 ordinary hidden layers plus one num_nextn_predict_layers setting. The MLX checkpoint sanitizer drops layers at or beyond num_hidden_layers, and the safetensors headers contain layers 0-46 only." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8127495265573422, "kv_store_format": "mlx_bf16_latent_mla", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mlx_bf16_latent_mla", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-6bit-affine-glm4-moe-lite-latent-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and small F32 router correction-bias tensors. Dequantization, activation traffic, router compute, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 6-bit affine MLX quantization with group_size 64 and BF16 runtime dtype. weight_bytes_per_param records resident stored bytes divided by the safetensors index total_parameters value; exact resident, fixed, routed, and compressed-state byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio GLM 4.7 Flash MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.7-Flash-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 28387239288ecde2acd4096526d6a67b3ff388f2, the API records a public non-gated MIT text-generation repo with transformers, safetensors, glm4_moe_lite, mlx, conversational, endpoints_compatible, 6-bit, region:us, base_model zai-org/GLM-4.7-Flash, 238894 downloads, and safetensors storage parameters BF16 941818624, U32 5613207552, and F32 2944. The card describes this as a 6-bit MLX quantized version of GLM-4.7-Flash optimized for Apple Silicon." }, { "label": "LM Studio GLM 4.7 Flash MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-6bit/raw/28387239288ecde2acd4096526d6a67b3ff388f2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_lora_rank", "qk_rope_head_dim", "v_head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Glm4MoeLiteForCausalLM, glm4_moe_lite, bfloat16 runtime dtype, 6-bit affine MLX quantization with group_size 64, 47 hidden layers, one next-token-prediction layer setting, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10240, moe_intermediate_size 1536, 20 attention heads, 20 key/value heads, q_lora_rank 768, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, v_head_dim 256, 64 routed experts, 4 experts per token, 1 shared expert, tie_word_embeddings false, vocab size 154880, and 202752 max position embeddings." }, { "label": "GLM-4.7-Flash BF16 source config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison found no differences across 34 checked architecture fields: architectures, model_type, dtype, hidden/layer/MoE sizes, attention head counts, MLA dimensions, routing settings, tied embeddings, max positions, vocab size, norm epsilon, activation, RoPE theta, and attention bias/dropout. The same fields also match the audited 8-bit MLX sibling. The source repo already has an audited BF16 profile." }, { "label": "LM Studio GLM 4.7 Flash MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-6bit/raw/28387239288ecde2acd4096526d6a67b3ff388f2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 24336479232 bytes, total_parameters 29943393920, five shards, and 1970 tensors. Range-read safetensors headers found 24.336479232 GB of tensor payload: 22.452830208 GB U32, 1.883637248 GB BF16, and 0.000011776 GB F32. Ordinary text resident tensors, defined as layers 0-46 plus model.norm and lm_head, sum to 24.058934272 GB. model.embed_tokens contributes 0.277544960 GB resident-only, and no model.layers.47 tensors were present. Packed routed expert tensors under mlp.switch_mlp sum to 22.573744128 GB across layers 1-46 and divide by the leading 64-expert dimension into 0.352714752 GB per expert. Fixed ordinary text traffic, including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head, sums to 1.485190144 GB." }, { "label": "MLX-LM GLM4 MoE Lite implementation", "url": "https://github.com/ml-explore/mlx-lm/blob/2ed22318cd6a2fcc5c2e0caa1e1fb0ddeb7cafd5/mlx_lm/models/glm4_moe_lite.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found Glm4MoeLiteModel instantiates range(config.num_hidden_layers), so ordinary decode uses 47 layers. The sanitizer drops any model.layers index greater than or equal to num_hidden_layers. Glm4MoeLiteAttention computes compressed_kv, splits kv_latent and k_pe, and calls cache.update_and_fetch(kv_latent, k_pe), so MLX stores the latent 512-dimensional state plus 64-dimensional RoPE key state rather than expanded full K/V heads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned BF16 source config comparison, audited 8-bit MLX sibling comparison, model card, safetensors index, direct five-shard safetensors header byte grouping, and MLX-LM runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 6-bit MoE weights and used expanded full-KV cache dimensions. It is an ordinary cached text-decode memory-side profile for the MLX artifact and intentionally uses MLX's latent MLA cache path." }, { "id": "lmstudio-community--glm-4-7-flash-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/GLM-4.7-Flash-MLX-8bit", "title": "LM Studio GLM 4.7 Flash MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized GLM-4.7-Flash artifact.", "model_family": "glm4-moe-lite", "base_model_proof": { "base_model": "zai-org/GLM-4.7-Flash", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records zai-org/GLM-4.7-Flash as its base model. Manual comparison found no differences across 34 checked architecture fields between the MLX config and the audited BF16 source config. The MLX repo adds 8-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not the GLM4 MoE Lite architecture." }, "architecture": { "canonical_architecture_id": "glm-4-7-flash", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.820755968, "main_resident_weight_gb": 31.483737088, "auxiliary_resident_weight_gb": 0.33701888, "fixed_weight_gb": 1.964225536, "routed_expert_weight_gb": 0.461242368, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16_f32", "traffic_scope": "ordinary MLX text decode through model.layers.0-46, model.norm, and lm_head, excluding model.embed_tokens resident-only lookup tensors", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each ordinary generated text token; the MLX artifact has no model.layers.47 tensors despite the config declaring one next-token-prediction layer", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package stores MLX U32 packed tensors, BF16 quantization metadata, and tiny F32 router correction biases. Routed experts are packed under mlp.switch_mlp with a leading 64-expert dimension; switch_mlp tensors sum to 29.519511552 GB, or 0.461242368 GB per expert group across layers 1-46." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.054144, "notes": "MLX GLM4 MoE Lite latent-cache coefficient: 47 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.054144, "notes": "Decode reads the same cached kv_latent plus RoPE key state per active context token in this v1 memory-traffic approximation." }, "notes": "The audited MLX implementation caches kv_latent with rank 512 and k_pe with RoPE dimension 64 through cache.update_and_fetch, rather than storing expanded full K/V heads. This profile is runtime-specific to the MLX artifact; the BF16 Transformers profile remains expanded-K/V." }, "notes": "The served config records 47 ordinary hidden layers plus one num_nextn_predict_layers setting. The MLX checkpoint sanitizer drops layers at or beyond num_hidden_layers, and the safetensors headers contain layers 0-46 only." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.062697036047943, "kv_store_format": "mlx_bf16_latent_mla", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mlx_bf16_latent_mla", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-8bit-affine-glm4-moe-lite-latent-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and small F32 router correction-bias tensors. Dequantization, activation traffic, router compute, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an 8-bit affine MLX quantization with group_size 64 and BF16 runtime dtype. weight_bytes_per_param records resident stored bytes divided by the safetensors index total_parameters value; exact resident, fixed, routed, and compressed-state byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio GLM 4.7 Flash MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.7-Flash-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 0eca07a0c2944693945f21a18cf663722e3186b9, the API records a public non-gated MIT text-generation repo with transformers, safetensors, glm4_moe_lite, mlx, conversational, endpoints_compatible, 8-bit, region:us, base_model zai-org/GLM-4.7-Flash, 272241 downloads, and safetensors storage parameters BF16 941818624, U32 7484276736, and F32 2944. The card describes this as an 8-bit MLX quantized version of GLM-4.7-Flash optimized for Apple Silicon." }, { "label": "LM Studio GLM 4.7 Flash MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-8bit/raw/0eca07a0c2944693945f21a18cf663722e3186b9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_lora_rank", "qk_rope_head_dim", "v_head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Glm4MoeLiteForCausalLM, glm4_moe_lite, bfloat16 runtime dtype, 8-bit affine MLX quantization with group_size 64, 47 hidden layers, one next-token-prediction layer setting, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10240, moe_intermediate_size 1536, 20 attention heads, 20 key/value heads, q_lora_rank 768, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, v_head_dim 256, 64 routed experts, 4 experts per token, 1 shared expert, tie_word_embeddings false, vocab size 154880, and 202752 max position embeddings." }, { "label": "GLM-4.7-Flash BF16 source config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison found no differences across 34 checked architecture fields: architectures, model_type, dtype, hidden/layer/MoE sizes, attention head counts, MLA dimensions, routing settings, tied embeddings, max positions, vocab size, norm epsilon, activation, RoPE theta, and attention bias/dropout. The source repo already has an audited BF16 profile." }, { "label": "LM Studio GLM 4.7 Flash MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-8bit/raw/0eca07a0c2944693945f21a18cf663722e3186b9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 31820755968 bytes, total_parameters 29943393920, seven shards, and 1970 tensors. Range-read safetensors headers found 31.820755968 GB of tensor payload: 29.937106944 GB U32, 1.883637248 GB BF16, and 0.000011776 GB F32. Ordinary text resident tensors, defined as layers 0-46 plus model.norm and lm_head, sum to 31.483737088 GB. model.embed_tokens contributes 0.337018880 GB resident-only, and no model.layers.47 tensors were present. Packed routed expert tensors under mlp.switch_mlp sum to 29.519511552 GB across layers 1-46 and divide by the leading 64-expert dimension into 0.461242368 GB per expert. Fixed ordinary text traffic, including attention, dense layer 0 MLP, routers, shared experts, norms, and lm_head, sums to 1.964225536 GB." }, { "label": "MLX-LM GLM4 MoE Lite implementation", "url": "https://github.com/ml-explore/mlx-lm/blob/2ed22318cd6a2fcc5c2e0caa1e1fb0ddeb7cafd5/mlx_lm/models/glm4_moe_lite.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found Glm4MoeLiteModel instantiates range(config.num_hidden_layers), so ordinary decode uses 47 layers. The sanitizer drops any model.layers index greater than or equal to num_hidden_layers. Glm4MoeLiteAttention computes compressed_kv, splits kv_latent and k_pe, and calls cache.update_and_fetch(kv_latent, k_pe), so MLX stores the latent 512-dimensional state plus 64-dimensional RoPE key state rather than expanded full K/V heads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned BF16 source config comparison, model card, safetensors index, direct seven-shard safetensors header byte grouping, and MLX-LM runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit MoE weights and used expanded full-KV cache dimensions. It is an ordinary cached text-decode memory-side profile for the MLX artifact and intentionally uses MLX's latent MLA cache path." }, { "id": "lmstudio-community--llama-3-2-3b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Llama-3.2-3B-Instruct-GGUF", "title": "LM Studio Llama 3.2 3B Instruct GGUF Q3_K_L", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q3_K_L GGUF artifact of Llama 3.2 3B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-3B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, and GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.2-3B-Instruct as the quantized base. The base raw config is gated in this audit environment, so this profile uses the selected public GGUF header as the direct architecture source instead of copying the base config." }, "architecture": { "canonical_architecture_id": "llama-3-2-3b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.212749888, "swept_params_b": 3.212749888, "resident_weight_gb": 1.815347488, "swept_weight_gb": 1.80750976, "auxiliary_resident_weight_gb": 0.007837728, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.2-3B-Instruct-Q3_K_L.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected Q3_K_L linked file is 1.815347488 GB. Header tensor spans total 1.807509760 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007837728 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 28-layer decoder with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q3_K_L GGUF artifact. It uses the public GGUF metadata for architecture because the Meta base config is gated." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5650447595626841, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q3-k-l-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact uses a mixed Q3_K_L layout with Q3_K, Q5_K, Q6_K, and small F32 tensors. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "LM Studio Llama 3.2 3B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Llama-3.2-3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit c91307b5cf18c8106b1f8a6218c26ae4dbfee472, the API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.2-3B-Instruct, Llama 3.2 license, region:us, 251450 downloads, GGUF architecture llama, 131072 context length, gguf.total 3212749888, and gguf.totalFileSize 1815347488." }, { "label": "LM Studio Llama 3.2 3B Instruct GGUF model card", "url": "https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "quantization_recipe" ], "notes": "The card metadata records this as a Transformers/GGUF quantized derivative of meta-llama/Llama-3.2-3B-Instruct, quantized by bartowski for LM Studio distribution." }, { "label": "Llama 3.2 3B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 0cb88a4f764b7a12671c53f0838cd831a0843b95, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 3212749824 parameters." }, { "label": "Llama 3.2 3B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A direct raw config request returned HTTP 401 with GatedRepo. This GGUF profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "LM Studio Llama 3.2 3B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF/tree/c91307b5cf18c8106b1f8a6218c26ae4dbfee472", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks of all GGUF siblings found Q3_K_L 1.815347488 GB, Q4_K_M 2.019377440 GB, Q6_K 2.643853600 GB, and Q8_0 3.421899040 GB. The selected Q3_K_L artifact exactly matches API gguf.totalFileSize." }, { "label": "LM Studio Llama 3.2 3B Instruct Q3_K_L GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF/resolve/c91307b5cf18c8106b1f8a6218c26ae4dbfee472/Llama-3.2-3B-Instruct-Q3_K_L.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 31 metadata entries and 255 tensors. The linked file is 1.815347488 GB. Tensor spans sum to 1.807509760 GB: token_embd.weight 0.323205120 GB, blk.* tensors 1.484292096 GB, output_norm.weight 0.000012288 GB, and rope_freqs.weight 0.000000256 GB. Metadata/tokenizer/header/file overhead accounts for 0.007837728 GB. Stored tensor bytes split into Q3_K 0.756940800 GB, Q5_K 0.726663168 GB, Q6_K 0.323205120 GB, and F32 0.000700672 GB. The header records llama.block_count 28, context_length 131072, embedding_length 3072, feed_forward_length 8192, attention.head_count 24, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 500000, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card metadata, base-model API metadata, gated base-config access check, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected Q3_K_L artifact." }, "notes": "Use this profile for the API-selected LM Studio Llama 3.2 3B Instruct Q3_K_L GGUF artifact. Do not infer the gated base config directly; the architecture evidence is the selected GGUF header metadata." }, { "id": "lmstudio-community--qwen2-5-coder-14b-instruct-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit", "title": "LM Studio Qwen2.5 Coder 14B Instruct MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen2.5 Coder 14B Instruct artifact.", "model_family": "qwen2.5-coder-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-14B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, audited BF16 base profile, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen2.5-Coder-14B-Instruct as its base model. Manual comparison found matching bounds-relevant Qwen2ForCausalLM geometry, context, dtype, tied-embedding, and sliding-window fields versus the audited BF16 base profile. The MLX config adds quantization metadata and sets max_window_layers to 70 instead of 48, but use_sliding_window is false in both configs, so this field does not affect the profiled KV adapter." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 8.309352448, "swept_weight_gb": 7.871408128, "auxiliary_resident_weight_gb": 0.43794432, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "model tensors excluding model.embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 scales, biases, norms, and tiny unquantized side tensors. Logical parameter counts treat U32 weight elements as eight 4-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from the two shard headers and are authoritative for the bound." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served MLX config records use_sliding_window false, sliding_window 131072, 48 layers, 8 KV heads, 128 head dimension, BF16 runtime dtype, and 32768 max positions. Bounds Engine v1 charges full-context BF16 K and V streams for ordinary cached text decode." }, "notes": "Dense Qwen2ForCausalLM MLX profile using the served LM Studio config and tensor headers rather than deriving structure from the model name." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5625818218852775, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records 4-bit affine MLX quantization with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 4275f1d1fd379c8a5e8cc655c5a57ef03b912a29, the API records a public Apache-2.0 text-generation repo with mlx, safetensors, qwen2, code, 4-bit, and region:us tags, base_model Qwen/Qwen2.5-Coder-14B-Instruct, 150668 downloads, and safetensors parameters F16 462377984, U32 1846149120, total 2308527104 storage elements. The model card describes this as a 4-bit MLX quantization by bartowski optimized for Apple Silicon." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit/raw/4275f1d1fd379c8a5e8cc655c5a57ef03b912a29/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings false, 4-bit affine MLX quantization with group_size 64, bfloat16 runtime dtype, 48 layers, hidden size 5120, intermediate size 13824, 40 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, use_sliding_window false, sliding_window 131072, max_window_layers 70, use_cache true, and vocab size 152064." }, { "label": "Qwen2.5 Coder 14B Instruct BF16 base config/profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/aedcc2d42b622764e023cf882b6652e646b95671/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout", "kv_adapter" ], "notes": "Manual comparison against the audited BF16 base profile found matching bounds-relevant tensor geometry, context, sliding-window, vocabulary, dtype, and tied-embedding fields. The base profile records 14.770033664B logical resident parameters, 13.991465984B ordinary swept logical parameters, and a separate 0.778567680B input embedding tensor. The only checked difference outside quantization metadata is max_window_layers 70 in the MLX config versus 48 in the BF16 base; both configs set use_sliding_window false." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 4-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit/blob/4275f1d1fd379c8a5e8cc655c5a57ef03b912a29/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "context_notes" ], "notes": "The model card identifies Qwen/Qwen2.5-Coder-14B-Instruct as the original model, records MLX quantization by bartowski from mlx-examples, and describes 128K long-context support through YaRN rope scaling. The served config remains at 32768 max positions, so this profile uses the served default context." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit/raw/4275f1d1fd379c8a5e8cc655c5a57ef03b912a29/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index records total_size 8309352448 bytes across two shards. Direct safetensors header range reads found 1255 tensors totaling 8.309352448 GB: F16 0.924755968 GB and U32 7.384596480 GB. The linked objects total 8.309494233 GB including 0.000141785 GB safetensors header/container overhead; bounds use tensor spans. Logical reconstruction from U32 packed weights and unquantized F16 model tensors excluding MLX .scales/.biases gives 14.770033664B resident parameters. model.embed_tokens contributes 0.778567680B logical params and 0.437944320 GB resident-only. lm_head contributes 0.778567680B logical params and 0.437944320 GB swept. Other model-body tensors contribute 13.212898304B logical params and 7.433463808 GB swept. Ordinary text swept traffic is model excluding embed_tokens plus lm_head, totaling 13.991465984B logical params and 7.871408128 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned served MLX config, audited BF16 base profile/config comparison, model card, safetensors index, and direct two-shard safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and missed MLX scales, biases, BF16 side tensors, untied lm_head storage, and resident-only embedding bytes. It is an ordinary cached text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen2-5-coder-14b-instruct-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit", "title": "LM Studio Qwen2.5 Coder 14B Instruct MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen2.5 Coder 14B Instruct artifact.", "model_family": "qwen2.5-coder-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-14B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, audited BF16 base profile, audited MLX 4-bit sibling profile, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen2.5-Coder-14B-Instruct as its base model. Manual comparison found matching bounds-relevant Qwen2ForCausalLM geometry, context, dtype, tied-embedding, and sliding-window fields versus the audited BF16 base profile and MLX 4-bit sibling. The MLX config adds 8-bit quantization metadata and sets max_window_layers to 70 instead of 48, but use_sliding_window is false, so this field does not affect the profiled KV adapter." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 15.693948928, "swept_weight_gb": 14.866720768, "auxiliary_resident_weight_gb": 0.82722816, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "model tensors excluding model.embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens tensors are resident for token lookup but are not swept as full matrices for each generated text token", "notes": "MLX 8-bit safetensors report packed U32 storage elements plus BF16 scales, biases, norms, and tiny unquantized side tensors. Logical parameter counts treat U32 weight elements as four 8-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from the three shard headers and are authoritative for the bound." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served MLX config records use_sliding_window false, sliding_window 131072, 48 layers, 8 KV heads, 128 head dimension, BF16 runtime dtype, and 32768 max positions. Bounds Engine v1 charges full-context BF16 K and V streams for ordinary cached text decode." }, "notes": "Dense Qwen2ForCausalLM MLX profile using the served LM Studio config and tensor headers rather than deriving structure from the model name." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.062553362099094, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records 8-bit affine MLX quantization with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 8-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 6fec0bda9a7de3b76be0395d35fa9045cc9d4c68, the API records a public Apache-2.0 text-generation repo with mlx, safetensors, qwen2, code, 8-bit, and region:us tags, base_model Qwen/Qwen2.5-Coder-14B-Instruct, 118924 downloads, and safetensors parameters F16 462377984, U32 3692298240, total 4154676224 storage elements." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit/raw/6fec0bda9a7de3b76be0395d35fa9045cc9d4c68/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings false, 8-bit affine MLX quantization with group_size 64, bfloat16 runtime dtype, 48 layers, hidden size 5120, intermediate size 13824, 40 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, use_sliding_window false, sliding_window 131072, max_window_layers 70, use_cache true, and vocab size 152064." }, { "label": "Qwen2.5 Coder 14B Instruct BF16 base config/profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/aedcc2d42b622764e023cf882b6652e646b95671/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout", "kv_adapter" ], "notes": "Manual comparison against the audited BF16 base profile and audited MLX 4-bit sibling found matching bounds-relevant tensor geometry, context, sliding-window, vocabulary, dtype, and tied-embedding fields. The base profile records 14.770033664B logical resident parameters, 13.991465984B ordinary swept logical parameters, and a separate 0.778567680B input embedding tensor. The checked difference outside quantization metadata is max_window_layers 70 in the MLX config versus 48 in the BF16 base; both configs set use_sliding_window false." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 8-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit/blob/6fec0bda9a7de3b76be0395d35fa9045cc9d4c68/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "context_notes" ], "notes": "The model card identifies Qwen/Qwen2.5-Coder-14B-Instruct as the original model, records MLX quantization by bartowski from mlx-examples, and describes 128K long-context support through YaRN rope scaling. The served config remains at 32768 max positions, so this profile uses the served default context." }, { "label": "LM Studio Qwen2.5 Coder 14B Instruct MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit/raw/6fec0bda9a7de3b76be0395d35fa9045cc9d4c68/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index records total_size 15693948928 bytes across three shards. Direct shard HEAD checks record 15694091111 linked bytes, including 0.000142183 GB safetensors header/container overhead; bounds use tensor spans. Direct safetensors header range reads found 1255 tensors totaling 15.693948928 GB: F16 0.924755968 GB and U32 14.769192960 GB. Logical reconstruction from U32 packed weights and unquantized F16 model tensors excluding MLX .scales/.biases gives 14.770033664B resident parameters. model.embed_tokens contributes 0.778567680B logical params and 0.827228160 GB resident-only. lm_head contributes 0.778567680B logical params and 0.827228160 GB swept. Other model-body tensors contribute 13.212898304B logical params and 14.039492608 GB swept. Ordinary text swept traffic is model excluding embed_tokens plus lm_head, totaling 13.991465984B logical params and 14.866720768 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned served MLX config, audited BF16 base profile/config comparison, audited MLX 4-bit sibling, model card, safetensors index, and direct three-shard safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as flat 8-bit dense weights and missed MLX scales, biases, BF16 side tensors, untied lm_head storage, and resident-only embedding bytes. It is an ordinary cached text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-5-9b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.5-9B-GGUF", "title": "LM Studio Qwen3.5 9B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected LM Studio Q4_K_M GGUF artifact of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-9B. The selected GGUF header records the same Qwen3.5 9B text geometry as the Qwen config. The LM Studio GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.953803264, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.01711872, "resident_weight_gb": 5.627044256, "swept_weight_gb": 5.04394752, "auxiliary_resident_weight_gb": 0.583096736, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-9B-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected main Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode; mmproj-Qwen3.5-9B-BF16.gguf is a separate sidecar and is not included unless explicitly loaded for another workload", "notes": "The HF API gguf.totalFileSize matches Qwen3.5-9B-Q4_K_M.gguf, so this profile targets the Q4_K_M artifact. Header tensor spans total 5.616076800 GB, while the linked file size is 5.627044256 GB. The main GGUF contains token_embd.weight, blk.* tensors, output.weight, and output_norm.weight. It has no mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 layers with every fourth layer using full attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact. The separate BF16 multimodal projector sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6284529702170593, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the LM Studio Q4_K_M GGUF. Tensor spans split into Q4_K 3.546021888 GB, Q5_K 0.553648128 GB, Q6_K 1.512161280 GB, and F32 0.004245504 GB. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "LM Studio Qwen3.5 9B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 1379f25c6b505a3fc737bd7818cb09389cf807c1 records base_model Qwen/Qwen3.5-9B, Apache-2.0 license, region:us, 391812 downloads, GGUF architecture qwen35, 262144 context length, gguf.total 8953803264, and gguf.totalFileSize 5627044256." }, { "label": "LM Studio Qwen3.5 9B GGUF model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, base_model Qwen/Qwen3.5-9B, and says LM Studio produced the GGUF quantization using llama.cpp release b8185." }, { "label": "Qwen3.5 9B config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "LM Studio Qwen3.5 9B GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF/tree/1379f25c6b505a3fc737bd7818cb09389cf807c1", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3.5-9B-Q4_K_M.gguf is 5627044256 bytes, exactly matching API gguf.totalFileSize. Sibling files are Q6_K 7359259040 bytes, Q8_0 9527501216 bytes, and mmproj-Qwen3.5-9B-BF16.gguf 921704480 bytes. The selected profile excludes the separate mmproj sidecar unless a workload explicitly loads it." }, { "label": "LM Studio Qwen3.5 9B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF/resolve/1379f25c6b505a3fc737bd7818cb09389cf807c1/Qwen3.5-9B-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 34 metadata entries and 427 tensors. The linked file is 5.627044256 GB. Tensor spans sum to 5.616076800 GB: token_embd.weight 0.572129280 GB, blk.* tensors 4.209575936 GB, output.weight 0.834355200 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.010967456 GB. Tensor spans split into Q4_K 3.546021888 GB, Q5_K 0.553648128 GB, Q6_K 1.512161280 GB, and F32 0.004245504 GB. The header records qwen35.block_count 32, context_length 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no mmproj/vision/audio/MTP tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless the separate mmproj GGUF file is explicitly loaded by the workload." }, { "id": "lmstudio-community--qwen3-5-9b-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.5-9B-MLX-4bit", "title": "LM Studio Qwen3.5 9B MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3.5 9B artifact.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.5-9B as its base model. Manual comparison found matching root, text, and vision architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16/F32 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.409813744, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.4731292, "resident_weight_gb": 5.95006256, "swept_weight_gb": 4.46591232, "auxiliary_resident_weight_gb": 1.48415024, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16/F32 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 scales, BF16 biases, F32 recurrent scalars, and unquantized BF16 side tensors. Logical parameter counts treat U32 weight elements as eight 4-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 32 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.6323252215054683, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and tiny F32 recurrent scalar tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 4-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.5 9B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit b455506b0f574c74616dbcd56879bde38fafcff3, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 4-bit, and region:us tags, base_model Qwen/Qwen3.5-9B, 762722 downloads, and safetensors parameters BF16 736844272, U32 1119092736, F32 768, total 1855937776 storage elements. The card describes this as a 4-bit MLX quantized version of Qwen3.5-9B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.5 9B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-4bit/raw/b455506b0f574c74616dbcd56879bde38fafcff3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 4-bit affine MLX quantization with group_size 64, BF16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 9B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in 37 checked root, text, and vision architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base architecture." }, { "label": "LM Studio Qwen3.5 9B MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-4bit/raw/b455506b0f574c74616dbcd56879bde38fafcff3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 5950062560 bytes across two shards. Range-read safetensors headers found 1260 tensors totaling 5.950062560 GB: BF16 1.473688544 GB, U32 4.476370944 GB, and F32 0.000003072 GB. Logical reconstruction from U32 packed weights and unquantized BF16/F32 model tensors gives 9.409813744B resident parameters. language_model.model.embed_tokens contributes 1.017118720B logical params and 0.572129280 GB resident-only. language_model.lm_head contributes 1.017118720B logical params and 0.572129280 GB swept. Other language_model tensors contribute 6.919565824B logical params and 3.893783040 GB swept. vision_tower contributes 0.456010480B logical params and 0.912020960 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 7.936684544B logical params and 4.465912320 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and missed MLX scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-5-9b-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.5-9B-MLX-8bit", "title": "LM Studio Qwen3.5 9B MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3.5 9B artifact.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.5-9B as its base model. Manual comparison found matching root, text, and vision architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16/F32 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.409813744, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.4731292, "resident_weight_gb": 10.426433504, "swept_weight_gb": 8.433723904, "auxiliary_resident_weight_gb": 1.9927096, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16/F32 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 scales, BF16 biases, F32 recurrent scalars, and unquantized BF16 side tensors. Logical parameter counts treat U32 weight elements as four 8-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 32 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.1080382447153356, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and tiny F32 recurrent scalar tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an 8-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.5 9B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 950e0459b483aec4324437bc5d45cf446b2d67bd, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 8-bit, and region:us tags, base_model Qwen/Qwen3.5-9B, 668158 downloads, and safetensors parameters BF16 736844272, U32 2238185472, F32 768, total 2975030512 storage elements. The card describes this as an 8-bit MLX quantized version of Qwen3.5-9B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.5 9B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-8bit/raw/950e0459b483aec4324437bc5d45cf446b2d67bd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 8-bit affine MLX quantization with group_size 64, BF16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 9B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in 34 checked root, text, and vision architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base architecture." }, { "label": "LM Studio Qwen3.5 9B MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-8bit/raw/950e0459b483aec4324437bc5d45cf446b2d67bd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 10426433504 bytes across two shards. Range-read safetensors headers found 1260 tensors totaling 10.426433504 GB: BF16 1.473688544 GB, U32 8.952741888 GB, and F32 0.000003072 GB. Logical reconstruction from U32 packed weights and unquantized BF16/F32 model tensors gives 9.409813744B resident parameters. language_model.model.embed_tokens contributes 1.017118720B logical params and 1.080688640 GB resident-only. language_model.lm_head contributes 1.017118720B logical params and 1.080688640 GB swept. Other language_model tensors contribute 6.919565824B logical params and 7.353035264 GB swept. vision_tower contributes 0.456010480B logical params and 0.912020960 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 7.936684544B logical params and 8.433723904 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit dense weights and missed MLX scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-27b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-27B-GGUF", "title": "LM Studio Qwen3.6 27B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected LM Studio Q4_K_M GGUF artifact of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, pinned Qwen base config, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The LM Studio card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-27B. The selected Q4_K_M GGUF header records the same Qwen3.6 text geometry as the Qwen config, with 64 ordinary text layers and no MTP, mmproj, vision, or draft tensors in the main file." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 16.547398784, "swept_weight_gb": 15.821244416, "auxiliary_resident_weight_gb": 0.726154368, "resident_parameter_scope": "selected Qwen3.6-27B-Q4_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj sidecar is not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.6-27B-Q4_K_M.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. The selected linked file is 16.547398784 GB. Header tensor spans total 16.536406016 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010992768 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, and ordinary blk.0-63 tensors, with no MTP, mmproj, vision, or draft tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate mmproj-BF16 GGUF sidecar requires a separate workload profile if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6152364563133252, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.6-27B-Q4_K_M.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "LM Studio Qwen3.6 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 58c6607d9c4cae8b071b3781c73be633fb3dee36 records a public Apache-2.0 GGUF repo with base_model Qwen/Qwen3.6-27B, 331042 downloads, region:us, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 16547398784. The API totalFileSize exactly matches Qwen3.6-27B-Q4_K_M.gguf, so this profile targets that artifact." }, { "label": "LM Studio Qwen3.6 27B GGUF model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-GGUF/raw/58c6607d9c4cae8b071b3781c73be633fb3dee36/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "weight_format" ], "notes": "The card records Apache-2.0 licensing, base_model Qwen/Qwen3.6-27B, original model Qwen/Qwen3.6-27B, and GGUF quantization by the LM Studio team using llama.cpp release b8883." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP training settings." }, { "label": "LM Studio Qwen3.6 27B GGUF linked-object HEAD checks", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-GGUF/tree/58c6607d9c4cae8b071b3781c73be633fb3dee36", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwen3.6-27B-Q4_K_M.gguf 16.547398784 GB, Qwen3.6-27B-Q6_K.gguf 22.082528384 GB, Qwen3.6-27B-Q8_0.gguf 28.595762304 GB, and mmproj-Qwen3.6-27B-BF16.gguf 0.931145856 GB. Q4_K_M exactly matches API gguf.totalFileSize and the mmproj sidecar is not part of the selected main text artifact." }, { "label": "LM Studio Qwen3.6 27B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-GGUF/resolve/58c6607d9c4cae8b071b3781c73be633fb3dee36/Qwen3.6-27B-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 38 metadata entries and 851 tensors. The linked file is 16.547398784 GB. Tensor spans sum to 16.536406016 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.715161600 GB, and ordinary blk.0-63 tensors 14.778279936 GB. Metadata/tokenizer/header/file overhead accounts for 0.010992768 GB. Tensor spans split into Q4_K 12.076646400 GB, Q6_K 4.449177600 GB, and F32 0.010582016 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, and no MTP, mmproj, vision, or draft tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned Qwen base config, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the LM Studio main Q4_K_M GGUF text artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or MTP speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "lmstudio-community--qwen3-6-27b-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-27B-MLX-4bit", "title": "LM Studio Qwen3.6 27B MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3.6 27B artifact.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-27B as its base model. Manual comparison found matching root and text architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 16.05426224, "swept_weight_gb": 14.417640448, "auxiliary_resident_weight_gb": 1.636621792, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as eight 4-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 64 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5868487602524941, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 4-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 27B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit bd83f6fe15b171f1549475db2348389c0f541c21, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 4-bit, and region:us tags, base_model Qwen/Qwen3.6-27B, 1132490 downloads, and safetensors parameters BF16 1303792880 plus U32 3361669120, total 4665462000 storage elements. The card describes this as a 4-bit MLX quantized version of Qwen3.6-27B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 27B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-4bit/raw/bd83f6fe15b171f1549475db2348389c0f541c21/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, 4-bit affine MLX quantization with group_size 64, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the checked root and text architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base text architecture." }, { "label": "LM Studio Qwen3.6 27B MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-4bit/raw/bd83f6fe15b171f1549475db2348389c0f541c21/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 16054262240 bytes across three shards. Range-read safetensors headers found 2180 tensors totaling 16.05426224 GB: BF16 2.60758576 GB and U32 13.44667648 GB. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors gives 27.35672856B resident parameters. language_model.model.embed_tokens contributes 1.2713984B logical params and 0.7151616 GB resident-only. language_model.lm_head contributes 1.2713984B logical params and 0.7151616 GB swept. Other language_model tensors contribute 24.353201664B logical params and 13.702478848 GB swept. vision_tower contributes 0.460730096B logical params and 0.921460192 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 25.624600064B logical params and 14.417640448 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and missed MLX scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-27b-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-27B-MLX-5bit", "title": "LM Studio Qwen3.6 27B MLX 5-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 5-bit quantized Qwen3.6 27B artifact.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-27B as its base model. Manual comparison found matching root and text architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 19.41593136, "swept_weight_gb": 17.620384768, "auxiliary_resident_weight_gb": 1.795546592, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as 32/5 packed 5-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 64 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.7097314767522773, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-5bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 5-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 27B MLX 5-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-5bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 589f789ab2a6673feaf18e2acaa0867b07f657d8, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 5-bit, and region:us tags, base_model Qwen/Qwen3.6-27B, 1065162 downloads, and safetensors parameters BF16 1303792880 plus U32 4202086400. The card describes this as a 5-bit MLX quantized version of Qwen3.6-27B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 27B MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-5bit/raw/589f789ab2a6673feaf18e2acaa0867b07f657d8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, 5-bit affine MLX quantization with group_size 64, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the checked root and text architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base text architecture." }, { "label": "LM Studio Qwen3.6 27B MLX 5-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-5bit/raw/589f789ab2a6673feaf18e2acaa0867b07f657d8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 19415931360 bytes across four shards. Range-read safetensors headers found 2180 tensors totaling 19.41593136 GB: BF16 2.60758576 GB and U32 16.8083456 GB. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors gives 27.35672856B resident parameters. language_model.model.embed_tokens contributes 1.2713984B logical params and 0.8740864 GB resident-only. language_model.lm_head contributes 1.2713984B logical params and 0.8740864 GB swept. Other language_model tensors contribute 24.353201664B logical params and 16.746298368 GB swept. vision_tower contributes 0.460730096B logical params and 0.921460192 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 25.624600064B logical params and 17.620384768 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense weights and missed MLX 5-bit packing, scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-27b-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-27B-MLX-6bit", "title": "LM Studio Qwen3.6 27B MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Qwen3.6 27B artifact.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-27B as its base model. Manual comparison found matching root and text architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 22.77760048, "swept_weight_gb": 20.823129088, "auxiliary_resident_weight_gb": 1.954471392, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as 32/6 packed 6-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 64 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8326141932520604, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 6-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 27B MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit dafc96bab96a89e3ea8c4b3c197e2806b8c0d595, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 6-bit, and region:us tags, base_model Qwen/Qwen3.6-27B, 1070949 downloads, and safetensors parameters BF16 1303792880 plus U32 5042503680, total 6346296560 storage elements. The card describes this as a 6-bit MLX quantized version of Qwen3.6-27B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 27B MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-6bit/raw/dafc96bab96a89e3ea8c4b3c197e2806b8c0d595/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, 6-bit affine MLX quantization with group_size 64, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the checked root and text architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base text architecture." }, { "label": "LM Studio Qwen3.6 27B MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-6bit/raw/dafc96bab96a89e3ea8c4b3c197e2806b8c0d595/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 22777600480 bytes across five shards. Range-read safetensors headers found 2180 tensors totaling 22.77760048 GB: BF16 2.60758576 GB and U32 20.17001472 GB. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors gives 27.35672856B resident parameters. language_model.model.embed_tokens contributes 1.2713984B logical params and 1.0330112 GB resident-only. language_model.lm_head contributes 1.2713984B logical params and 1.0330112 GB swept. Other language_model tensors contribute 24.353201664B logical params and 19.790117888 GB swept. vision_tower contributes 0.460730096B logical params and 0.921460192 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 25.624600064B logical params and 20.823129088 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as BF16/U32 dense weights and missed MLX 6-bit packing, scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-27b-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-27B-MLX-8bit", "title": "LM Studio Qwen3.6 27B MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3.6 27B artifact.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-27B as its base model. Manual comparison found matching root and text architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 29.50093872, "swept_weight_gb": 27.228617728, "auxiliary_resident_weight_gb": 2.272320992, "resident_parameter_scope": "logical model parameters reconstructed from MLX packed U32 weights plus unquantized BF16 model tensors", "swept_parameter_scope": "language_model tensors excluding embed_tokens plus language_model.lm_head safetensors headers", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "MLX safetensors report packed U32 storage elements plus BF16 quantization metadata and unquantized side tensors. Logical parameter counts treat U32 weight elements as four 8-bit parameters and exclude MLX .scales/.biases metadata. Stored-byte fields are exact sums from shard headers and are authoritative for the bound. The headers contain 64 language layers and no mtp-named tensors, so this package's resident logical count is lower than the full BF16 base package that also stores MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.0783796262516265, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and unquantized side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an 8-bit affine MLX quantization_config with group_size 64 and BF16 text runtime settings. weight_bytes_per_param records resident stored bytes divided by the reconstructed resident logical parameter count; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 27B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 75c2bb5358faad8cea31345d1112d79544a3d311, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 8-bit, and region:us tags, base_model Qwen/Qwen3.6-27B, 1144822 downloads, and safetensors parameters BF16 1303792880 plus U32 6723338240, total 8027131120 storage elements. The card describes this as an 8-bit MLX quantized version of Qwen3.6-27B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 27B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-8bit/raw/75c2bb5358faad8cea31345d1112d79544a3d311/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, 8-bit affine MLX quantization with group_size 64, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the checked root and text architecture fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base text architecture." }, { "label": "LM Studio Qwen3.6 27B MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-8bit/raw/75c2bb5358faad8cea31345d1112d79544a3d311/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "The index records total_size 29500938720 bytes across six shards. Range-read safetensors headers found 2180 tensors totaling 29.50093872 GB: BF16 2.60758576 GB and U32 26.89335296 GB. Logical reconstruction from U32 packed weights and unquantized BF16 model tensors gives 27.35672856B resident parameters. language_model.model.embed_tokens contributes 1.2713984B logical params and 1.3508608 GB resident-only. language_model.lm_head contributes 1.2713984B logical params and 1.3508608 GB swept. Other language_model tensors contribute 24.353201664B logical params and 25.877756928 GB swept. vision_tower contributes 0.460730096B logical params and 0.921460192 GB resident-only. Ordinary text swept traffic is language_model excluding embed_tokens plus lm_head, totaling 25.624600064B logical params and 27.228617728 GB. The indexed headers contain no mtp-named tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit dense weights and missed MLX scales, biases, BF16 side tensors, vision tensors, untied lm_head storage, and the fixed DeltaNet runtime state. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-35B-A3B-GGUF", "title": "LM Studio Qwen3.6 35B A3B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected LM Studio Q4_K_M GGUF artifact of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-35B-A3B. The selected Q4_K_M GGUF header records the same audited Qwen3.6 text geometry as the Qwen config: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 routed experts per token, and a separate always-on shared expert." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 21.166757728, "main_resident_weight_gb": 20.869704192, "auxiliary_resident_weight_gb": 0.297053536, "fixed_weight_gb": 1.366190592, "routed_expert_weight_gb": 0.0761856, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3.6-35B-A3B-Q4_K_M.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; the separate mmproj-Qwen3.6-35B-A3B-BF16.gguf sidecar is not included unless explicitly loaded for another workload", "shared_expert_notes": "The Qwen model card states 8 routed plus 1 shared expert, and the GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected main GGUF mixes Q4_K, Q6_K, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and speculative paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6106862316574311, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative execution are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.6-35B-A3B-Q4_K_M.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "LM Studio Qwen3.6 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 68a34855558af61cbef0324d31f411be8a506b08 records a public Apache-2.0 GGUF repo with base_model Qwen/Qwen3.6-35B-A3B, 308995 downloads, region:us, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 21166757728. The API totalFileSize matches Qwen3.6-35B-A3B-Q4_K_M.gguf, so this profile targets that artifact." }, { "label": "LM Studio Qwen3.6 35B A3B GGUF model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-GGUF/raw/68a34855558af61cbef0324d31f411be8a506b08/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact", "quantization_tool" ], "notes": "The card records Apache-2.0 licensing, base_model Qwen/Qwen3.6-35B-A3B, original model Qwen/Qwen3.6-35B-A3B, and GGUF quantization by the LM Studio team using llama.cpp release b8814." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The current base config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, and 262144 max position embeddings." }, { "label": "LM Studio Qwen3.6 35B A3B Q4_K_M linked object and GGUF range-read tensor index", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-GGUF/resolve/68a34855558af61cbef0324d31f411be8a506b08/Qwen3.6-35B-A3B-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 41 metadata entries and 733 tensors. The linked file is 21.166757728 GB. Tensor spans sum to 21.155768832 GB; metadata/tokenizer/header/file overhead accounts for 0.010988896 GB. Tensor spans split into Q4_K 16.030679040 GB, Q6_K 5.036236800 GB, and F32 0.088852992 GB. token_embd.weight is 0.286064640 GB and resident-only; output.weight is swept. Routed expert tensors sum to 19.503513600 GB, or 0.076185600 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 1.366190592 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found Q4_K_M 21.166757728 GB, Q6_K 28.514152288 GB, Q8_0 36.903139168 GB, and separate mmproj BF16 0.902822016 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, current Qwen base config, selected linked file size, a direct GGUF header/tensor-index range read of the API-selected Q4_K_M artifact, sidecar HEAD checks, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected LM Studio Qwen3.6 35B A3B Q4_K_M main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "lmstudio-community--qwen3-6-35b-a3b-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit", "title": "LM Studio Qwen3.6 35B A3B MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3.6 35B A3B artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-35B-A3B as its base model. Manual comparison found matching root, text, and vision architecture fields between the MLX config and the Qwen base config. The MLX repo adds quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 20.401929952, "main_resident_weight_gb": 19.222722816, "auxiliary_resident_weight_gb": 1.179207136, "fixed_weight_gb": 1.103329536, "routed_expert_weight_gb": 0.07077888, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16", "traffic_scope": "ordinary text decode through language_model excluding embed_tokens plus language_model.lm_head, with packed switch_mlp tensors divided by the leading 256-expert dimension", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the multimodal package but not swept as full matrices for each ordinary generated text token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package stores MLX U32 packed tensors and BF16 side tensors. Routed experts are packed under mlp.switch_mlp with a leading 256-expert dimension; switch_mlp tensors sum to 18.11939328 GB, or 0.07077888 GB per expert group across all 40 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. MLX weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and config-level MTP settings. This MLX artifact's safetensors headers contain no MTP-named tensors, so the profile models ordinary text decode after any multimodal prefill with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5674797108334115, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and recurrent side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 4-bit affine MLX quantization_config with group_size 64 and per-layer gate/shared_expert_gate overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the base logical BF16 parameter count; exact resident, fixed, and routed byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 35B A3B MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b, the API records an Apache-2.0 image-text-to-text repo with mlx, endpoints_compatible, 4-bit, and region:us tags, base_model Qwen/Qwen3.6-35B-A3B, 648427 downloads, and safetensors storage parameters BF16 1530838896, U32 4335063040, total 5865901936. The card describes this as a 4-bit MLX quantized version of Qwen3.6-35B-A3B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit/raw/0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, 4-bit affine MLX quantization with group_size 64 and 8-bit overrides for every layer's gate and shared_expert_gate modules, BF16 text config, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 35B A3B base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The base card states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited root, text_config, and vision_config geometry fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base architecture." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit/raw/0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 20401929952 bytes across four shards. Range-read safetensors headers found 2090 tensors totaling 20.401929952 GB: 17.34025216 GB U32 and 3.061677792 GB BF16. Ordinary text resident tensors, defined as language_model excluding model.embed_tokens plus lm_head, sum to 19.222722816 GB. Auxiliary resident tensors, defined as vision_tower plus language_model.model.embed_tokens, sum to 1.179207136 GB. The input embedding contributes 0.28606464 GB, the vision tower contributes 0.893142496 GB, and no MTP-named tensors or layer-40 tensors were present. Packed routed expert tensors under mlp.switch_mlp sum to 18.11939328 GB across all layers and divide by the leading 256-expert dimension into 0.07077888 GB per expert. Fixed ordinary text traffic, including routers, shared experts, shared expert gates, attention, DeltaNet weights, norms, and lm_head, sums to 1.103329536 GB. The lm_head contributes 0.28606464 GB and the input embedding contributes the same resident-only storage." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights and missed MLX scales, biases, BF16 side tensors, vision tensors, shared-expert traffic, per-session DeltaNet runtime state, and the packed switch_mlp expert layout. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-35b-a3b-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit", "title": "LM Studio Qwen3.6 35B A3B MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Qwen3.6 35B A3B artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-35B-A3B as its base model. Manual comparison found matching root, text, and vision architecture fields between the MLX config and the Qwen base config. The MLX repo adds 6-bit affine quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 29.061529312, "main_resident_weight_gb": 27.755182336, "auxiliary_resident_weight_gb": 1.306346976, "fixed_weight_gb": 3.595991296, "routed_expert_weight_gb": 0.09437184, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16", "traffic_scope": "ordinary text decode through language_model excluding embed_tokens plus language_model.lm_head, with packed switch_mlp tensors divided by the leading 256-expert dimension", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors, including MLX embedding scales and biases, are resident for the multimodal package but not swept as full matrices for each ordinary generated text token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package stores MLX U32 packed tensors and BF16 side tensors. Routed experts are packed under mlp.switch_mlp with a leading 256-expert dimension; switch_mlp tensors sum to 24.15919104 GB, or 0.09437184 GB per expert group across all 40 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. MLX weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and config-level MTP settings. This MLX artifact's safetensors headers contain no MTP-named tensors, so the profile models ordinary text decode after any multimodal prefill with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8083469899142742, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and recurrent side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a 6-bit affine MLX quantization_config with group_size 64 and per-layer gate/shared_expert_gate overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the base logical BF16 parameter count; exact resident, fixed, and routed byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 35B A3B MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit ee1b3d2dc76e36e305a48f24b168a903fec6bc0d, the API records an Apache-2.0 image-text-to-text repo with transformers, safetensors, qwen3_5_moe, mlx, endpoints_compatible, 6-bit, and region:us tags, base_model Qwen/Qwen3.6-35B-A3B, 618161 downloads, and safetensors storage parameters BF16 1530838896, U32 6499962880, total 8030801776. The card describes this as a 6-bit MLX quantized version of Qwen3.6-35B-A3B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit/raw/ee1b3d2dc76e36e305a48f24b168a903fec6bc0d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, 6-bit affine MLX quantization with group_size 64 and 8-bit overrides for every layer's gate and shared_expert_gate modules, BF16 text config, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 35B A3B base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The base card states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited root, text_config, and vision_config geometry fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base architecture." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit/raw/ee1b3d2dc76e36e305a48f24b168a903fec6bc0d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 29061529312 bytes across six shards. Range-read safetensors headers found 2090 tensors totaling 29.061529312 GB: 25.99985152 GB U32 and 3.061677792 GB BF16. Ordinary text resident tensors, defined as language_model excluding model.embed_tokens plus language_model.lm_head, sum to 27.755182336 GB. Auxiliary resident tensors, defined as vision_tower plus all language_model.model.embed_tokens tensors, sum to 1.306346976 GB. The input embedding tensors contribute 0.41320448 GB, the vision tower contributes 0.893142496 GB, and no MTP-named tensors or layer-40 tensors were present. Packed routed expert tensors under mlp.switch_mlp sum to 24.15919104 GB across all layers and divide by the leading 256-expert dimension into 0.09437184 GB per expert. Fixed ordinary text traffic, including routers, shared experts, shared expert gates, attention, DeltaNet weights, norms, and all lm_head tensors, sums to 3.595991296 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as an ideal BF16/U32 MoE package and missed exact MLX scales, biases, BF16 side tensors, vision tensors, shared-expert traffic, per-session DeltaNet runtime state, and the packed switch_mlp expert layout. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-6-35b-a3b-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit", "title": "LM Studio Qwen3.6 35B A3B MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3.6 35B A3B artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3.6-35B-A3B as its base model. Manual comparison found matching root, text, and vision architecture fields between the MLX config and the Qwen base config. The MLX repo adds 8-bit affine quantization_config and packed U32/BF16 storage tensors, which change serving bytes but not the text architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 37.721128672, "main_resident_weight_gb": 36.287641856, "auxiliary_resident_weight_gb": 1.433486816, "fixed_weight_gb": 2.062121216, "routed_expert_weight_gb": 0.13369344, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16", "traffic_scope": "ordinary text decode through language_model excluding embed_tokens plus language_model.lm_head, with packed switch_mlp tensors divided by the leading 256-expert dimension", "auxiliary_scope": "vision_tower tensors and language_model.model.embed_tokens tensors are resident for the multimodal package but not swept as full matrices for each ordinary generated text token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package stores MLX U32 packed tensors and BF16 side tensors. Routed experts are packed under mlp.switch_mlp with a leading 256-expert dimension; switch_mlp tensors sum to 34.22552064 GB, or 0.13369344 GB per expert group across all 40 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. MLX weight quantization does not change the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and config-level MTP settings. This MLX artifact's safetensors headers contain no MTP-named tensors, so the profile models ordinary text decode after any multimodal prefill with speculative MTP disabled." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.0492139106247809, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weight tensors plus BF16 scales, biases, norms, vision tensors, and recurrent side tensors. Dequantization, activation traffic, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records an 8-bit affine MLX quantization_config with group_size 64 and per-layer gate/shared_expert_gate overrides also at 8-bit. weight_bytes_per_param records resident stored bytes divided by the base logical BF16 parameter count; exact resident, fixed, and routed byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3.6 35B A3B MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit ea4ab1544e989b4fa38711453e0296cd7b03ac43, the API records an Apache-2.0 image-text-to-text repo with transformers, safetensors, qwen3_5_moe, mlx, endpoints_compatible, 8-bit, and region:us tags, base_model Qwen/Qwen3.6-35B-A3B, 629948 downloads, and safetensors storage parameters BF16 1530838896, U32 8664862720, total 10195701616. The card describes this as an 8-bit MLX quantized version of Qwen3.6-35B-A3B optimized for Apple Silicon." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit/raw/ea4ab1544e989b4fa38711453e0296cd7b03ac43/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, 8-bit affine MLX quantization with group_size 64, BF16 text config, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 35B A3B base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The base card states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited root, text_config, and vision_config geometry fields between the base BF16 repo and this MLX artifact; the MLX artifact adds quantization_config while preserving the base architecture." }, { "label": "LM Studio Qwen3.6 35B A3B MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit/raw/ea4ab1544e989b4fa38711453e0296cd7b03ac43/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 37721128672 bytes across eight shards. Range-read safetensors headers found 2090 tensors totaling 37.721128672 GB: 34.65945088 GB U32 and 3.061677792 GB BF16. Ordinary text resident tensors, defined as language_model excluding model.embed_tokens plus lm_head, sum to 36.287641856 GB. Auxiliary resident tensors, defined as vision_tower plus language_model.model.embed_tokens, sum to 1.433486816 GB. The input embedding contributes 0.54034432 GB, the vision tower contributes 0.893142496 GB, and no MTP-named tensors or layer-40 tensors were present. Packed routed expert tensors under mlp.switch_mlp sum to 34.22552064 GB across all layers and divide by the leading 256-expert dimension into 0.13369344 GB per expert. Fixed ordinary text traffic, including routers, shared experts, shared expert gates, attention, DeltaNet weights, norms, and lm_head, sums to 2.062121216 GB. The lm_head contributes 0.54034432 GB and the input embedding contributes the same resident-only storage." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served MLX config, pinned base Qwen config comparison, model card, safetensors index, direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit MoE weights and missed MLX scales, biases, BF16 side tensors, vision tensors, shared-expert traffic, per-session DeltaNet runtime state, and the packed switch_mlp expert layout. It is an ordinary text-decode memory-side profile for the MLX artifact." }, { "id": "lmstudio-community--qwen3-coder-30b-a3b-instruct-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit", "title": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3-Coder 30B A3B Instruct artifact.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": false, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the memory-relevant ordinary text-decode topology: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, disabled sliding-window attention, and 262144 max positions. Manual comparison against the official BF16 base config found served-config differences in intermediate_size, shared_expert_intermediate_size, and max_window_layers, so this profile uses the MLX served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-mlx-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.174622208, "main_resident_weight_gb": 16.999591936, "auxiliary_resident_weight_gb": 0.175030272, "fixed_weight_gb": 0.692137984, "routed_expert_weight_gb": 0.127401984, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct packed switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but are not swept as a full matrix for each generated text token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived stored bytes are authoritative for the bound. MLX safetensors store packed U32 weights plus BF16 scales, biases, norms, and side tensors. Routed expert tensors are packed under model.layers.*.mlp.switch_mlp.* with a leading 128-expert dimension; switch_mlp tensors sum to 16.307453952 GB, or 0.127401984 GB per expert group across all 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM MLX profile using the served LM Studio config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, MLX dequantization, activation traffic, and Apple MLX runtime scheduling are outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5625099315728466, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-4bit-affine-qwen3-coder-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head, attention tensors, and router tensors. Dequantization, activation traffic, kernel choice, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with 4-bit affine group_size 64. weight_bytes_per_param records resident stored bytes divided by the API/index total parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 4-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit bbecedefd826819d2c6b6465e88ca7b9b8ad3407, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_moe, mlx, endpoints_compatible, region:us, 4-bit, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 183045. The API safetensors summary records BF16 954333184 and U32 3816488960 storage elements with total 30532122624 parameters." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 4-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit/raw/bbecedefd826819d2c6b6465e88ca7b9b8ad3407/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, says the artifact is a 4-bit quantized version using MLX, and states that it is optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit/raw/bbecedefd826819d2c6b6465e88ca7b9b8ad3407/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, and 4-bit affine MLX quantization with group_size 64." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked memory-relevant architecture fields. The LM Studio MLX config matches the official BF16 base on layer count, attention heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, shared_expert_intermediate_size 0 vs null, and max_window_layers 28 vs 48, so config_compatible is false and fixed_weight_gb is derived from the served MLX tensor headers." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit/raw/bbecedefd826819d2c6b6465e88ca7b9b8ad3407/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 17174622208 bytes across four shards and total_parameters 30532122624. Direct range-read safetensors headers found 1351 tensors with payload bytes exactly matching the index total: 17.174622208 GB, split into U32 15.265955840 GB and BF16 1.908666368 GB. model.embed_tokens U32 weight plus BF16 scales/biases total 0.175030272 GB resident-only for ordinary decode. lm_head is stored separately with the same 0.175030272 GB layout and remains in fixed decode traffic. Packed switch_mlp routed expert tensors sum to 16.307453952 GB and divide by the leading 128-expert dimension into 0.127401984 GB per expert group. Non-expert fixed decode tensors including lm_head sum to 0.692137984 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated ideal 0.5-byte metadata estimate for this MLX repo. It uses exact stored MLX tensor payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "lmstudio-community--qwen3-coder-30b-a3b-instruct-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit", "title": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 5-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 5-bit quantized Qwen3-Coder 30B A3B Instruct artifact.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": false, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the memory-relevant ordinary text-decode topology: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, disabled sliding-window attention, and 262144 max positions. Manual comparison against the official BF16 base config found served-config differences in intermediate_size, shared_expert_intermediate_size, and max_window_layers, so this profile uses the MLX served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-mlx-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 20.991111168, "main_resident_weight_gb": 20.77718528, "auxiliary_resident_weight_gb": 0.213925888, "fixed_weight_gb": 0.845852672, "routed_expert_weight_gb": 0.155713536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct packed switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but are not swept as a full matrix for each generated text token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived stored bytes are authoritative for the bound. MLX safetensors store packed U32 weights plus BF16 scales, biases, norms, and side tensors. Routed expert tensors are packed under model.layers.*.mlp.switch_mlp.* with a leading 128-expert dimension; switch_mlp tensors sum to 19.931332608 GB, or 0.155713536 GB per expert group across all 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM MLX profile using the served LM Studio config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, MLX dequantization, activation traffic, and Apple MLX runtime scheduling are outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.6875090679578164, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-5bit-affine-qwen3-coder-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head, attention tensors, and router tensors. Dequantization, activation traffic, kernel choice, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with 5-bit affine group_size 64. weight_bytes_per_param records resident stored bytes divided by the API/index total parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 5-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit a7d91f62488598d4725c5363a73d0a8f0e1b9329, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_moe, mlx, endpoints_compatible, region:us, 5-bit, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 160568. The API safetensors summary records BF16 954333184 and U32 4770611200 storage elements with total 30532122624 parameters." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 5-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit/raw/a7d91f62488598d4725c5363a73d0a8f0e1b9329/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, says the artifact is a 5-bit quantized version using MLX, and states that it is optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit/raw/a7d91f62488598d4725c5363a73d0a8f0e1b9329/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, and 5-bit affine MLX quantization with group_size 64." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked memory-relevant architecture fields. The LM Studio MLX config matches the official BF16 base on layer count, attention heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, shared_expert_intermediate_size 0 vs null, and max_window_layers 28 vs 48, so config_compatible is false and fixed_weight_gb is derived from the served MLX tensor headers." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 5-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit/raw/a7d91f62488598d4725c5363a73d0a8f0e1b9329/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 20991111168 bytes across four shards and total_parameters 30532122624. Direct range-read safetensors headers found 1351 tensors with payload bytes exactly matching the index total: 20.991111168 GB, split into U32 19.082444800 GB and BF16 1.908666368 GB. model.embed_tokens U32 weight plus BF16 scales/biases total 0.213925888 GB resident-only for ordinary decode. lm_head is stored separately with the same 0.213925888 GB layout and remains in fixed decode traffic. Packed switch_mlp routed expert tensors sum to 19.931332608 GB and divide by the leading 128-expert dimension into 0.155713536 GB per expert group. Non-expert fixed decode tensors including lm_head sum to 0.845852672 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated ideal 0.625-byte metadata estimate for this MLX repo. It uses exact stored MLX tensor payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "lmstudio-community--qwen3-coder-30b-a3b-instruct-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit", "title": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Qwen3-Coder 30B A3B Instruct artifact.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": false, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the memory-relevant ordinary text-decode topology: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, disabled sliding-window attention, and 262144 max positions. Manual comparison against the official BF16 base config found served-config differences in intermediate_size, shared_expert_intermediate_size, and max_window_layers, so this profile uses the MLX served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-mlx-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.807600128, "main_resident_weight_gb": 24.554778624, "auxiliary_resident_weight_gb": 0.252821504, "fixed_weight_gb": 0.99956736, "routed_expert_weight_gb": 0.184025088, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct packed switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but are not swept as a full matrix for each generated text token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived stored bytes are authoritative for the bound. MLX safetensors store packed U32 weights plus BF16 scales, biases, norms, and side tensors. Routed expert tensors are packed under model.layers.*.mlp.switch_mlp.* with a leading 128-expert dimension; switch_mlp tensors sum to 23.555211264 GB, or 0.184025088 GB per expert group across all 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM MLX profile using the served LM Studio config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, MLX dequantization, activation traffic, and Apple MLX runtime scheduling are outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8125082043427864, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-6bit-affine-qwen3-coder-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head, attention tensors, and router tensors. Dequantization, activation traffic, kernel choice, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with 6-bit affine group_size 64. weight_bytes_per_param records resident stored bytes divided by the API/index total parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 6-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_moe, mlx, endpoints_compatible, region:us, 6-bit, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 159385. The API safetensors summary records BF16 954333184 and U32 5724733440 storage elements with total 30532122624 parameters." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 6-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit/raw/0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, says the artifact is a 6-bit quantized version using MLX, and states that it is optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit/raw/0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, and 6-bit affine MLX quantization with group_size 64." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked memory-relevant architecture fields. The LM Studio MLX config matches the official BF16 base on layer count, attention heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, shared_expert_intermediate_size 0 vs null, and max_window_layers 28 vs 48, so config_compatible is false and fixed_weight_gb is derived from the served MLX tensor headers." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit/raw/0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 24807600128 bytes across five shards and total_parameters 30532122624. Direct range-read safetensors headers found 1351 tensors with payload bytes exactly matching the index total: 24.807600128 GB, split into U32 22.898933760 GB and BF16 1.908666368 GB. model.embed_tokens U32 weight plus BF16 scales/biases total 0.252821504 GB resident-only for ordinary decode. lm_head is stored separately with the same 0.252821504 GB layout and remains in fixed decode traffic. Packed switch_mlp routed expert tensors sum to 23.555211264 GB and divide by the leading 128-expert dimension into 0.184025088 GB per expert group. Non-expert fixed decode tensors including lm_head sum to 0.999567360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated ideal 0.75-byte metadata estimate for this MLX repo. It uses exact stored MLX tensor payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "lmstudio-community--qwen3-coder-30b-a3b-instruct-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit", "title": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3-Coder 30B A3B Instruct artifact.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": false, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the memory-relevant ordinary text-decode topology: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, disabled sliding-window attention, and 262144 max positions. Manual comparison against the official BF16 base config found served-config differences in intermediate_size, shared_expert_intermediate_size, and max_window_layers, so this profile uses the MLX served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-mlx-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.440578048, "main_resident_weight_gb": 32.109965312, "auxiliary_resident_weight_gb": 0.330612736, "fixed_weight_gb": 1.306996736, "routed_expert_weight_gb": 0.240648192, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct packed switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but are not swept as a full matrix for each generated text token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived stored bytes are authoritative for the bound. MLX safetensors store packed U32 weights plus BF16 scales, biases, norms, and side tensors. Routed expert tensors are packed under model.layers.*.mlp.switch_mlp.* with a leading 128-expert dimension; switch_mlp tensors sum to 30.802968576 GB, or 0.240648192 GB per expert group across all 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM MLX profile using the served LM Studio config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, MLX dequantization, activation traffic, and Apple MLX runtime scheduling are outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.062506477112726, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-8bit-affine-qwen3-coder-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, lm_head, attention tensors, and router tensors. Dequantization, activation traffic, kernel choice, expert compute, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with 8-bit affine group_size 64. weight_bytes_per_param records resident stored bytes divided by the API/index total parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 8-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 4285def38c15e8a2a01bf136a1a2349484003540, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen3_moe, mlx, endpoints_compatible, region:us, 8-bit, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 163274. The API safetensors summary records BF16 954333184 and U32 7632977920 storage elements with total 30532122624 parameters." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 8-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit/raw/4285def38c15e8a2a01bf136a1a2349484003540/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, says the artifact is an 8-bit quantized version using MLX, and states that it is optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit/raw/4285def38c15e8a2a01bf136a1a2349484003540/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, and 8-bit affine MLX quantization with group_size 64." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked memory-relevant architecture fields. The LM Studio MLX config matches the official BF16 base on layer count, attention heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, shared_expert_intermediate_size 0 vs null, and max_window_layers 28 vs 48, so config_compatible is false and fixed_weight_gb is derived from the served MLX tensor headers." }, { "label": "LM Studio Qwen3-Coder 30B A3B Instruct MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit/raw/4285def38c15e8a2a01bf136a1a2349484003540/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 32440578048 bytes across seven shards and total_parameters 30532122624. Direct range-read safetensors headers found 1351 tensors with payload bytes exactly matching the index total: 32.440578048 GB, split into U32 30.531911680 GB and BF16 1.908666368 GB. model.embed_tokens U32 weight plus BF16 scales/biases total 0.330612736 GB resident-only for ordinary decode. lm_head is stored separately with the same 0.330612736 GB layout and remains in fixed decode traffic. Packed switch_mlp routed expert tensors sum to 30.802968576 GB and divide by the leading 128-expert dimension into 0.240648192 GB per expert group. Non-expert fixed decode tensors including lm_head sum to 1.306996736 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated ideal 1-byte metadata estimate for this MLX repo. It uses exact stored MLX tensor payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "lmstudio-community--qwen3-coder-next-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-Next-MLX-4bit", "title": "LM Studio Qwen3-Coder-Next MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3-Coder-Next artifact.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-Next as its base model. Manual comparison found no differences in checked memory-relevant architecture fields between the MLX config and the official BF16 base config. The MLX repo adds 4-bit affine quantization metadata, 8-bit router/shared-expert-gate overrides, and packed U32/BF16 storage tensors, which change serving bytes but not model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 44.84406016, "main_resident_weight_gb": 44.669029888, "auxiliary_resident_weight_gb": 0.175030272, "fixed_weight_gb": 1.182486016, "routed_expert_weight_gb": 0.084934656, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but not swept as a full matrix for each generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. Shared expert, router gate, and shared expert gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "MLX safetensors report packed U32 storage elements rather than base BF16 tensor bytes. Header-derived stored bytes are authoritative for the bound. Routed expert tensors are stored under model.layers.*.mlp.switch_mlp.* and are byte-uniform across all 48 layers and 512 expert indexes; switch_mlp tensors total 43.486543872 GB, exactly 0.905969664 GB per layer and 0.084934656 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5628415784615021, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-4bit-affine-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, embedding metadata, and recurrent scalar tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, router compute, expert compute, and state writes are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with default 4-bit affine group_size 64 and 96 per-layer gate/shared_expert_gate overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the reconstructed logical parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder-Next MLX 4-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 03cba26036330b7553d252c1f3fb899f16cc5ea5, the API records a public, non-gated Apache-2.0 MLX repo with safetensors, qwen3_next, base_model:Qwen/Qwen3-Coder-Next, 4-bit, region:us, and 181136 downloads. The API safetensors block reports BF16 2491172608 and U32 9965428736 storage elements. The card states that this is a 4-bit quantized version of Qwen3-Coder-Next using MLX, optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder-Next MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-4bit/raw/03cba26036330b7553d252c1f3fb899f16cc5ea5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and 4-bit affine MLX quantization with group_size 64 plus 8-bit overrides for every layer's gate and shared_expert_gate modules." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no checked architecture differences between the official BF16 base config and the MLX 4-bit config: architecture, model type, dtype, untied embeddings, layer count, full_attention_interval, hidden size, intermediate sizes, attention heads, KV heads, head dimensions, expert geometry, linear-attention state geometry, max positions, and vocabulary size match. The MLX config only adds quantization_config." }, { "label": "LM Studio Qwen3-Coder-Next MLX 4-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-4bit/raw/03cba26036330b7553d252c1f3fb899f16cc5ea5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 44844060160 bytes across nine shards. Direct range-read safetensors headers found 1891 tensors totaling the same 44.844060160 GB: BF16 4.982345216 GB and U32 39.861714944 GB. Reconstructed logical parameters from packed U32 weights, 8-bit router/shared-expert-gate overrides, and unquantized model tensors sum to 79.674391296B, matching the official BF16 base API; API/index total_parameters is lower by 2304 elements and is not used for bounds. Resident-only model.embed_tokens tensors total 0.175030272 GB. Main resident tensors total 44.669029888 GB. switch_mlp routed expert tensors total 43.486543872 GB, exactly 0.905969664 GB in each of 48 layers and 0.084934656 GB per expert index across 512 expert indexes. Fixed ordinary text traffic totals 1.182486016 GB, including linear attention, self attention, routers, shared expert, shared expert gates, norms, and lm_head tensors." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, official BF16 base config comparison, model card text, safetensors index, direct shard header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights and missed MLX scales, biases, 8-bit router overrides, shared experts, routed switch_mlp naming, exact stored bytes, and the fixed DeltaNet runtime state." }, { "id": "lmstudio-community--qwen3-coder-next-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-Next-MLX-6bit", "title": "LM Studio Qwen3-Coder-Next MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Qwen3-Coder-Next artifact.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-Next as its base model. Manual comparison found no differences in checked memory-relevant architecture fields between the MLX config and the official BF16 base config. The MLX repo adds 6-bit affine quantization metadata, 8-bit router/shared-expert-gate overrides, and packed U32/BF16 storage tensors, which change serving bytes but not model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 64.749702656, "main_resident_weight_gb": 64.496881152, "auxiliary_resident_weight_gb": 0.252821504, "fixed_weight_gb": 1.682984448, "routed_expert_weight_gb": 0.122683392, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but not swept as a full matrix for each generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. Shared expert, router gate, and shared expert gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "MLX safetensors report packed U32 storage elements rather than base BF16 tensor bytes. Header-derived stored bytes are authoritative for the bound. Routed expert tensors are stored under model.layers.*.mlp.switch_mlp.* and are byte-uniform across all 48 layers and 512 expert indexes; switch_mlp tensors total 62.813896704 GB, exactly 1.308622848 GB per layer and 0.122683392 GB per expert index." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8126789750479175, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-6bit-affine-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, embedding metadata, and recurrent scalar tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, router compute, expert compute, and state writes are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with default 6-bit affine group_size 64 and 96 per-layer gate/shared_expert_gate overrides at 8-bit. weight_bytes_per_param records resident stored bytes divided by the reconstructed logical parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder-Next MLX 6-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-6bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 6b4712e5519c7a7c5612992c5fca1608f42a14c2, the API records a public, non-gated Apache-2.0 MLX repo with safetensors, qwen3_next, base_model:Qwen/Qwen3-Coder-Next, 6-bit, region:us, and 179477 downloads. The API safetensors block reports BF16 2491172608 and U32 14941839360 storage elements." }, { "label": "LM Studio Qwen3-Coder-Next MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-6bit/raw/6b4712e5519c7a7c5612992c5fca1608f42a14c2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and 6-bit affine MLX quantization with group_size 64 plus 8-bit overrides for every layer's gate and shared_expert_gate modules." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no checked architecture differences between the official BF16 base config and the MLX 6-bit config across 24 memory-relevant fields: architecture, model type, dtype, untied embeddings, layer count, full_attention_interval, hidden size, intermediate sizes, attention heads, KV heads, head dimensions, expert geometry, linear-attention state geometry, max positions, sliding-window flag, and vocabulary size match. The MLX config only adds quantization_config." }, { "label": "Qwen3-Coder-Next BF16 base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-Next", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b" ], "notes": "At commit a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb, the official base API records Apache-2.0 text-generation metadata, region:us, and BF16 safetensors total 79674391296 parameters. This profile uses 79.674391296B as the logical parameter denominator; the MLX index metadata total_parameters is lower by 2304 elements and is not used for bounds." }, { "label": "LM Studio Qwen3-Coder-Next MLX 6-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-6bit/raw/6b4712e5519c7a7c5612992c5fca1608f42a14c2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 64749702656 bytes across 13 shards. Direct range-read safetensors headers found 1891 tensors totaling the same 64.749702656 GB: BF16 4.982345216 GB and U32 59.767357440 GB. Resident-only model.embed_tokens tensors total 0.252821504 GB. Main resident tensors total 64.496881152 GB. switch_mlp routed expert tensors total 62.813896704 GB, exactly 1.308622848 GB in each of 48 layers and 0.122683392 GB per expert index across 512 expert indexes. Fixed ordinary text traffic totals 1.682984448 GB, including linear attention, self attention, routers, shared expert, shared expert gates, norms, and lm_head tensors." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, official BF16 base config/API comparison, safetensors index, direct shard header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 6-bit MoE weights and missed MLX scales, biases, 8-bit router overrides, shared experts, routed switch_mlp naming, exact stored bytes, and the fixed DeltaNet runtime state." }, { "id": "lmstudio-community--qwen3-coder-next-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-Coder-Next-MLX-8bit", "title": "LM Studio Qwen3-Coder-Next MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3-Coder-Next artifact.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-Coder-Next as its base model. Manual comparison found no differences in checked memory-relevant architecture fields between the MLX config and the official BF16 base config. The MLX repo adds 8-bit affine quantization metadata and packed U32/BF16 storage tensors, which change serving bytes but not model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 84.655345152, "main_resident_weight_gb": 84.324732416, "auxiliary_resident_weight_gb": 0.330612736, "fixed_weight_gb": 2.18348288, "routed_expert_weight_gb": 0.160432128, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "MLX safetensors stored-byte total from model.safetensors.index.json and direct shard headers", "traffic_scope": "ordinary text decode excludes resident-only token embeddings and charges fixed language tensors plus expected distinct switch_mlp routed expert tensors", "auxiliary_scope": "model.embed_tokens U32 weight plus BF16 scales/biases are resident for token lookup but not swept as a full matrix for each generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. Shared expert, router gate, and shared expert gate tensors are outside switch_mlp and are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "MLX safetensors report packed U32 storage elements rather than base BF16 tensor bytes. Header-derived stored bytes are authoritative for the bound. Routed expert tensors are stored under model.layers.*.mlp.switch_mlp.*, not .experts.*, and are byte-uniform across all 48 layers and 512 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. Quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.0625163716343329, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-8bit-affine-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored MLX safetensors bytes: packed U32 expert and fixed-language weight tensors plus BF16 scales, biases, norms, embedding metadata, and recurrent scalar tensors. Dequantization, activation traffic, Apple MLX runtime scheduling, router compute, expert compute, and state writes are outside Bounds Engine v1.", "notes": "The config records bfloat16 runtime dtype plus an MLX quantization_config with default 8-bit affine group_size 64 and 96 per-layer gate/shared_expert_gate overrides also at 8-bit. weight_bytes_per_param records resident stored bytes divided by the reconstructed logical parameter denominator; exact resident/fixed/expert byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3-Coder-Next MLX 8-bit model card and API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-8bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At commit f911d6c6cb4cf62fd3c5fd4909feb046872e0221, the API records a public, non-gated Apache-2.0 MLX repo with safetensors, qwen3_next, base_model:Qwen/Qwen3-Coder-Next, 8-bit, region:us, and 203906 downloads. The API safetensors block reports BF16 2491172608 and U32 19918249984 storage elements. The card states that this is an 8-bit quantized version of Qwen3-Coder-Next using MLX, optimized for Apple Silicon." }, { "label": "LM Studio Qwen3-Coder-Next MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-8bit/raw/f911d6c6cb4cf62fd3c5fd4909feb046872e0221/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, and 8-bit affine MLX quantization with group_size 64." }, { "label": "Qwen3-Coder-Next BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_comparison" ], "notes": "Manual comparison found no checked architecture differences between the official BF16 base config and the MLX 8-bit config: architecture, model type, dtype, untied embeddings, layer count, full_attention_interval, hidden size, intermediate sizes, attention heads, KV heads, head dimensions, expert geometry, linear-attention state geometry, max positions, and vocabulary size match. The MLX config only adds quantization_config." }, { "label": "LM Studio Qwen3-Coder-Next MLX 8-bit safetensors index and shard headers", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-8bit/raw/f911d6c6cb4cf62fd3c5fd4909feb046872e0221/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 84655345152 bytes across 17 shards. Direct range-read safetensors headers found 1891 tensors totaling the same 84.655345152 GB: BF16 4.982345216 GB and U32 79.672999936 GB. Reconstructed logical parameters from packed U32 weights plus unquantized model tensors sum to 79.674391296B, matching the official BF16 base API; index metadata total_parameters is lower by 2304 elements and is not used for bounds. Resident-only model.embed_tokens tensors total 0.330612736 GB. Main resident tensors total 84.324732416 GB. switch_mlp routed expert tensors total 82.141249536 GB, exactly 1.711276032 GB in each of 48 layers and 0.160432128 GB per expert index across 512 expert indexes. Fixed ordinary text traffic totals 2.183482880 GB, including linear attention, self attention, routers, shared expert, shared expert gates, norms, and lm_head tensors." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served MLX config, official BF16 base config comparison, model card text, safetensors index, direct shard header byte grouping, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 8-bit MoE weights and missed MLX scales, biases, shared experts, routed switch_mlp naming, exact stored bytes, and the fixed DeltaNet runtime state." }, { "id": "lmstudio-community--qwen3-vl-4b-instruct-mlx-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit", "title": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 4-bit quantized Qwen3-VL 4B Instruct artifact.", "model_family": "qwen3-vl-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-VL-4B-Instruct as its quantized base. Manual comparison against the official BF16 base config found matching checked memory-relevant text and vision geometry: Qwen3VLForConditionalGeneration, qwen3_vl_text, 36 text layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max positions, tied text embeddings, and the 24-layer visual tower. The target repo adds MLX 4-bit affine quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.563511808, "swept_params_b": 4.148164096, "auxiliary_resident_params_b": 0.415347712, "resident_weight_gb": 3.093615616, "swept_weight_gb": 2.262920192, "auxiliary_resident_weight_gb": 0.830695424, "resident_parameter_scope": "direct MLX safetensors stored logical count with U32 packed weights counted as eight 4-bit weights and BF16 side tensors counted as stored resident elements", "swept_parameter_scope": "ordinary text decode includes all language_model.model tensors including tied embed_tokens; the header stores no lm_head tensor", "auxiliary_scope": "vision_tower tensors are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. model.safetensors stores packed U32 language weights plus BF16 scales, biases, embeddings, norms, and unquantized vision tensors. The checked-in model.safetensors.index.json is stale for this revision because it names two shards that are not present on the Hub; the actual downloadable file is a single model.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. It does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.677902401956489, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-qwen3-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 4-bit affine quantization with group_size 64. weight_bytes_per_param is the stored resident payload divided by the profile's logical stored-element resident denominator; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/revision/552af30c9952c44f1e1a27c7c5810ded58e892bc", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "storage_summary" ], "notes": "At commit 552af30c9952c44f1e1a27c7c5810ded58e892bc, the live API records a public non-gated Apache-2.0 image-text-to-text repo with mlx, safetensors, qwen3_vl, base_model Qwen/Qwen3-VL-4B-Instruct, 4-bit, and region:us tags. Current downloads are 110453. The API safetensors summary records BF16 541239808 storage elements, U32 502784000 storage elements, and total 1044023808 stored elements; direct headers are used for the logical packed-weight denominator and byte totals." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/raw/552af30c9952c44f1e1a27c7c5810ded58e892bc/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-VL-4B-Instruct base and says this package is a 4-bit quantized MLX artifact optimized for Apple Silicon." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/raw/552af30c9952c44f1e1a27c7c5810ded58e892bc/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Qwen3VLForConditionalGeneration, qwen3_vl_text, root and text tie_word_embeddings true, BF16 text dtype, 36 text layers, hidden_size 2560, intermediate_size 9728, 32 attention heads, 8 KV heads, head_dim 128, vocab_size 151936, 262144 max position embeddings, M-RoPE metadata, a resident 24-layer visual tower, and MLX 4-bit affine quantization with group_size 64." }, { "label": "Qwen3 VL 4B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in the checked memory-relevant architecture fields between the base BF16 config and the LM Studio MLX served config. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit safetensors header", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/resolve/552af30c9952c44f1e1a27c7c5810ded58e892bc/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Range-reading model.safetensors found format metadata mlx, a 151895-byte header, and 1219 tensors totaling 3.093615616 GB of payload bytes. Payload bytes split into U32 2.011136000 GB and BF16 1.082479616 GB. language_model.model tensors total 2.262920192 GB and are the ordinary swept text-decode traffic because the config ties embeddings and the header has no lm_head tensor. vision_tower tensors total 0.830695424 GB resident-only. Counting U32 packed tensors as eight 4-bit weights plus BF16 side tensors gives 4.563511808B resident logical stored elements, 4.148164096B swept logical stored elements, and 0.415347712B auxiliary resident logical stored elements." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 4-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/raw/552af30c9952c44f1e1a27c7c5810ded58e892bc/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The index metadata records total_size 8875631616 and maps 713 weights to model-00001-of-00002.safetensors and model-00002-of-00002.safetensors, but the Hub file listing and dry-run download at the pinned revision expose a single 3.1 GB model.safetensors file. Direct range requests for the indexed shard names return 404, so this profile treats the index as stale and uses the actual downloadable safetensors header." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI dry-run listing, model card, pinned served MLX config, pinned Qwen base config comparison, stale index review, and direct safetensors header byte grouping." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps the resident visual tower separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--qwen3-vl-4b-instruct-mlx-5bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit", "title": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 5-bit quantized Qwen3-VL 4B Instruct artifact.", "model_family": "qwen3-vl-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-VL-4B-Instruct as its quantized base. Manual comparison against the official BF16 base config found matching checked memory-relevant text and vision geometry: Qwen3VLForConditionalGeneration, qwen3_vl_text, 36 text layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max positions, tied text embeddings, and the 24-layer visual tower. The target repo adds MLX 5-bit affine quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.563511808, "swept_params_b": 4.148164096, "auxiliary_resident_params_b": 0.415347712, "resident_weight_gb": 3.596399616, "swept_weight_gb": 2.765704192, "auxiliary_resident_weight_gb": 0.830695424, "resident_parameter_scope": "direct MLX safetensors stored logical count with U32 packed tensor bits counted as 5-bit weights and BF16 side tensors counted as stored resident elements", "swept_parameter_scope": "ordinary text decode includes all language_model.model tensors including tied embed_tokens; the header stores no lm_head tensor", "auxiliary_scope": "vision_tower tensors are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. model.safetensors stores packed U32 language weights plus BF16 scales, biases, embeddings, norms, and unquantized vision tensors. The checked-in model.safetensors.index.json is stale for this revision because it names two shards that are not present on the Hub; the actual downloadable file is a single model.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. It does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.7880772017934482, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-5bit-affine-qwen3-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 5-bit affine quantization with group_size 64. weight_bytes_per_param is the stored resident payload divided by the profile's logical stored-element resident denominator; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/revision/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "storage_summary" ], "notes": "At commit 2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0, the live API records a public non-gated Apache-2.0 image-text-to-text repo with mlx, safetensors, qwen3_vl, base_model Qwen/Qwen3-VL-4B-Instruct, 5-bit, and region:us tags. Current downloads are 106372. The API safetensors summary records BF16 541239808 storage elements, U32 628480000 storage elements, and total 1169719808 stored elements; direct headers are used for the logical packed-weight denominator and byte totals." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/raw/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-VL-4B-Instruct base and says this package is a 5-bit quantized MLX artifact optimized for Apple Silicon." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/raw/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Qwen3VLForConditionalGeneration, qwen3_vl_text, root and text tie_word_embeddings true, BF16 text dtype, 36 text layers, hidden_size 2560, intermediate_size 9728, 32 attention heads, 8 KV heads, head_dim 128, vocab_size 151936, 262144 max position embeddings, M-RoPE metadata, a resident 24-layer visual tower, and MLX 5-bit affine quantization with group_size 64." }, { "label": "Qwen3 VL 4B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in the checked memory-relevant architecture fields between the base BF16 config and the LM Studio MLX served config. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit safetensors header", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/resolve/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Range-reading model.safetensors found format metadata mlx, a 152041-byte header, and 1219 tensors totaling 3.596399616 GB of payload bytes. Payload bytes split into U32 2.513920000 GB and BF16 1.082479616 GB. language_model.model tensors total 2.765704192 GB and are the ordinary swept text-decode traffic because the config ties embeddings and the header has no lm_head tensor. vision_tower tensors total 0.830695424 GB resident-only. Counting U32 packed tensor bits as 5-bit weights plus BF16 side tensors gives 4.563511808B resident logical stored elements, 4.148164096B swept logical stored elements, and 0.415347712B auxiliary resident logical stored elements." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 5-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/raw/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The index metadata records total_size 8875631616 and maps 713 weights to model-00001-of-00002.safetensors and model-00002-of-00002.safetensors, but the Hub file listing and dry-run download at the pinned revision expose a single 3.6 GB model.safetensors file. Direct HEAD requests for the indexed shard names return 404, so this profile treats the index as stale and uses the actual downloadable safetensors header." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI dry-run listing, model card, pinned served MLX config, pinned Qwen base config comparison, stale index review, and direct safetensors header byte grouping." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps the resident visual tower separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--qwen3-vl-4b-instruct-mlx-6bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit", "title": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 6-bit quantized Qwen3-VL 4B Instruct artifact.", "model_family": "qwen3-vl-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-VL-4B-Instruct as its quantized base. Manual comparison against the official BF16 base config found matching checked memory-relevant text and vision geometry: Qwen3VLForConditionalGeneration, qwen3_vl_text, 36 text layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max positions, tied text embeddings, and the 24-layer visual tower. The target repo adds MLX 6-bit affine quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.563511808, "swept_params_b": 4.148164096, "auxiliary_resident_params_b": 0.415347712, "resident_weight_gb": 4.099183616, "swept_weight_gb": 3.268488192, "auxiliary_resident_weight_gb": 0.830695424, "resident_parameter_scope": "direct MLX safetensors stored logical count with U32 packed tensor bits counted as 6-bit weights and BF16 side tensors counted as stored resident elements", "swept_parameter_scope": "ordinary text decode includes all language_model.model tensors including tied embed_tokens; the header stores no lm_head tensor", "auxiliary_scope": "vision_tower tensors are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. model.safetensors stores packed U32 language weights plus BF16 scales, biases, embeddings, norms, and unquantized vision tensors. The checked-in model.safetensors.index.json is stale for this revision because it names two shards that are not present on the Hub; the actual downloadable file is a single model.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. It does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.8982520016304075, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-6bit-affine-qwen3-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 6-bit affine quantization with group_size 64. weight_bytes_per_param is the stored resident payload divided by the profile's logical stored-element resident denominator; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/revision/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "storage_summary" ], "notes": "At commit 31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a, the live API records a public non-gated Apache-2.0 image-text-to-text repo with mlx, safetensors, qwen3_vl, base_model Qwen/Qwen3-VL-4B-Instruct, 6-bit, and region:us tags. Current downloads are 106558. The API safetensors summary records BF16 541239808 storage elements, U32 754176000 storage elements, and total 1295415808 stored elements; direct headers are used for the logical packed-weight denominator and byte totals." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/raw/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-VL-4B-Instruct base and says this package is a 6-bit quantized MLX artifact optimized for Apple Silicon." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/raw/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Qwen3VLForConditionalGeneration, qwen3_vl_text, root and text tie_word_embeddings true, BF16 text dtype, 36 text layers, hidden_size 2560, intermediate_size 9728, 32 attention heads, 8 KV heads, head_dim 128, vocab_size 151936, 262144 max position embeddings, M-RoPE metadata, a resident 24-layer visual tower, and MLX 6-bit affine quantization with group_size 64." }, { "label": "Qwen3 VL 4B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in the checked memory-relevant architecture fields between the base BF16 config and the LM Studio MLX served config. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit safetensors header", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/resolve/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Range-reading model.safetensors found format metadata mlx, a 152133-byte header, and 1219 tensors totaling 4.099183616 GB of payload bytes. Payload bytes split into U32 3.016704000 GB and BF16 1.082479616 GB. language_model.model tensors total 3.268488192 GB and are the ordinary swept text-decode traffic because the config ties embeddings and the header has no lm_head tensor. vision_tower tensors total 0.830695424 GB resident-only. Counting U32 packed tensor bits as 6-bit weights plus BF16 side tensors gives 4.563511808B resident logical stored elements, 4.148164096B swept logical stored elements, and 0.415347712B auxiliary resident logical stored elements." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 6-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/raw/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The index metadata records total_size 8875631616 and maps 713 weights to model-00001-of-00002.safetensors and model-00002-of-00002.safetensors, but the Hub file listing and dry-run download at the pinned revision expose a single 4.1 GB model.safetensors file. Direct HEAD requests for the indexed shard names return 404, so this profile treats the index as stale and uses the actual downloadable safetensors header." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI dry-run listing, model card, pinned served MLX config, pinned Qwen base config comparison, stale index review, and direct safetensors header byte grouping." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps the resident visual tower separate from ordinary text-decode traffic." }, { "id": "lmstudio-community--qwen3-vl-4b-instruct-mlx-8bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit", "title": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit", "summary": "Audited memory-side text-decode bounds profile for the LM Studio MLX 8-bit quantized Qwen3-VL 4B Instruct artifact.", "model_family": "qwen3-vl-dense-mlx", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served MLX config comparison, and direct MLX safetensors header grouping", "config_compatible": true, "notes": "The LM Studio repo records Qwen/Qwen3-VL-4B-Instruct as its quantized base. Manual comparison against the official BF16 base config found matching checked memory-relevant text and vision geometry: Qwen3VLForConditionalGeneration, qwen3_vl_text, 36 text layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max positions, tied text embeddings, and the 24-layer visual tower. The target repo adds MLX 8-bit affine quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.563511808, "swept_params_b": 4.148164096, "auxiliary_resident_params_b": 0.415347712, "resident_weight_gb": 5.104751616, "swept_weight_gb": 4.274056192, "auxiliary_resident_weight_gb": 0.830695424, "resident_parameter_scope": "direct MLX safetensors stored logical count with U32 packed weights counted as four 8-bit weights and BF16 side tensors counted as stored resident elements", "swept_parameter_scope": "ordinary text decode includes all language_model.model tensors including tied embed_tokens; the header stores no lm_head tensor", "auxiliary_scope": "vision_tower tensors are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "Exact range-read bytes drive the bound. model.safetensors stores packed U32 language weights plus BF16 scales, biases, embeddings, norms, and unquantized vision tensors. The checked-in model.safetensors.index.json is stale for this revision because it names two shards that are not present on the Hub; the actual downloadable file is a single model.safetensors." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. It does not define a sliding-window, recurrent-state text cache, or KV quantization scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary cached text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 1.1186016013043258, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-8bit-affine-qwen3-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, embeddings, and visual tensors. MLX dequantization, activation traffic, vision prefill, and Apple MLX runtime scheduling overhead are outside Bounds Engine v1.", "notes": "The config records a bfloat16 runtime dtype for text plus MLX 8-bit affine quantization with group_size 64. weight_bytes_per_param is the stored resident payload divided by the profile's logical stored-element resident denominator; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/revision/e73e3fbb7dae9f23283989a81a05d327a3958f3f", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "storage_summary" ], "notes": "At commit e73e3fbb7dae9f23283989a81a05d327a3958f3f, the live API records a public non-gated Apache-2.0 image-text-to-text repo with mlx, safetensors, qwen3_vl, base_model Qwen/Qwen3-VL-4B-Instruct, 8-bit, and region:us tags. Current downloads are 106978. The API safetensors summary records BF16 541239808 storage elements, U32 1005568000 storage elements, and total 1546807808 stored elements; direct headers are used for the logical packed-weight denominator and byte totals." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/raw/e73e3fbb7dae9f23283989a81a05d327a3958f3f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-VL-4B-Instruct base and says this package is an 8-bit quantized MLX artifact optimized for Apple Silicon." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/raw/e73e3fbb7dae9f23283989a81a05d327a3958f3f/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The served config records Qwen3VLForConditionalGeneration, qwen3_vl_text, root and text tie_word_embeddings true, BF16 text dtype, 36 text layers, hidden_size 2560, intermediate_size 9728, 32 attention heads, 8 KV heads, head_dim 128, vocab_size 151936, 262144 max position embeddings, M-RoPE metadata, a resident 24-layer visual tower, and MLX 8-bit affine quantization with group_size 64." }, { "label": "Qwen3 VL 4B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences in the checked memory-relevant architecture fields between the base BF16 config and the LM Studio MLX served config. The target repo only adds quantization and quantization_config blocks." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit safetensors header", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/resolve/e73e3fbb7dae9f23283989a81a05d327a3958f3f/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "Range-reading model.safetensors found format metadata mlx, a 152183-byte header, and 1219 tensors totaling 5.104751616 GB of payload bytes. Payload bytes split into U32 4.022272000 GB and BF16 1.082479616 GB. language_model.model tensors total 4.274056192 GB and are the ordinary swept text-decode traffic because the config ties embeddings and the header has no lm_head tensor. vision_tower tensors total 0.830695424 GB resident-only. Counting U32 packed tensors as four 8-bit weights plus BF16 side tensors gives 4.563511808B resident logical stored elements, 4.148164096B swept logical stored elements, and 0.415347712B auxiliary resident logical stored elements." }, { "label": "LM Studio Qwen3 VL 4B Instruct MLX 8-bit stale safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/raw/e73e3fbb7dae9f23283989a81a05d327a3958f3f/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "selected_artifact", "index_staleness" ], "notes": "The index metadata records total_size 8875631616 and maps 713 weights to model-00001-of-00002.safetensors and model-00002-of-00002.safetensors, but the Hub file listing and dry-run download at the pinned revision expose a single 5.1 GB model.safetensors file. Direct HEAD requests for the indexed shard names return 404, so this profile treats the index as stale and uses the actual downloadable safetensors header." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI dry-run listing, model card, pinned served MLX config, pinned Qwen base config comparison, stale index review, and direct safetensors header byte grouping." }, "notes": "This self-contained profile supersedes the scraped metadata estimate and keeps the resident visual tower separate from ordinary text-decode traffic." }, { "id": "lukealonso--minimax-m2-7-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lukealonso/MiniMax-M2.7-NVFP4", "title": "MiniMax M2.7 NVFP4", "summary": "Audited memory-side text-decode bounds profile for the ModelOpt NVFP4 MiniMax M2.7 MoE repo.", "model_family": "minimax-m2-moe-nvfp4", "base_model_proof": { "base_model": "MiniMaxAI/MiniMax-M2.7", "relation": "quantized", "source": "Hugging Face model card/API metadata, served ModelOpt config, hf_quant_config, audited FP8 base profile comparison, and direct safetensors header metadata", "config_compatible": false, "notes": "The API metadata and model card identify MiniMaxAI/MiniMax-M2.7 as the base model. Manual comparison against the audited FP8 base config found matching MiniMaxM2ForCausalLM tensor geometry, MoE routing geometry, rope settings, and full-context attention behavior. The served NVFP4 config records max_position_embeddings 196608 while the official FP8 base profile records 204800, so this profile uses the served NVFP4 config directly." }, "architecture": { "canonical_architecture_id": "minimax-m2-7", "max_context_tokens": 196608, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 134.401725368, "main_resident_weight_gb": 133.172532152, "auxiliary_resident_weight_gb": 1.229193216, "fixed_weight_gb": 6.789383168, "routed_expert_weight_gb": 0.49368417571875, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_modelopt_nvfp4_bf16_f8_f32_u8", "traffic_scope": "ordinary text decode through non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header data_offsets are used as the byte source of truth because the resolved safetensors index total_size includes header bytes. Expert tensors include ModelOpt NVFP4 U8 payloads, F8_E4M3 weight scales, F32 scale_2 tensors, and F32 input_scale tensors from the indexed model-inputscales sidecar. Per-expert bytes are nearly uniform; routed_expert_weight_gb is the exact routed expert byte sum divided by 256." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null, matching the audited FP8 base profile's full-context MiniMax M2 behavior. This profile charges full-context K and V streams for all 62 language layers." }, "notes": "MiniMaxM2ForCausalLM MoE profile using the served custom config, hf_quant_config, model card, audited base profile comparison, and direct safetensors header grouping." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.587667614175157, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "sglang-modelopt-nvfp4-moe-bf16-kv-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4/BF16/F8/F32/U8 safetensors bytes. NVFP4 dequantization, activation quantization, router compute, expert compute, PCIe all-reduce behavior, high-cache behavior, and writes are outside this memory-side bound.", "notes": "The model card's SGLang command uses --quantization modelopt_fp4 and --kv-cache-dtype bfloat16. The served config/hf_quant_config quantize Linear targets with 4-bit float weights and activations at group size 16 while ignoring lm_head, every MoE gate, and every self_attn module." }, "evidence": [ { "label": "MiniMax M2.7 NVFP4 API metadata", "url": "https://huggingface.co/api/models/lukealonso/MiniMax-M2.7-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "safetensors_dtype_split", "commit_sha" ], "notes": "At commit db821d7a3ce29ee96d80a1cae88d878d8586b54e, the API reports a public MIT safetensors repo with custom_code, minimax_m2, modelopt, 8-bit, region:us tags, base_model MiniMaxAI/MiniMax-M2.7, current downloads 518183, and safetensors stored elements BF16 4009288192, F8_E4M3 14042529792, U8 112340238336, total 130392056320." }, { "label": "MiniMax M2.7 NVFP4 model card", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "serving", "kv_adapter" ], "notes": "The card describes this as an NVFP4-quantized version of MiniMaxAI/MiniMax-M2.7. It states only MoE expert MLP gate/up/down projections are quantized to NVFP4, all other layers remain BF16, and the SGLang launch command uses --quantization modelopt_fp4, --kv-cache-dtype bfloat16, --context-length 196608, and tensor parallelism." }, { "label": "MiniMax M2.7 NVFP4 served config", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4/raw/db821d7a3ce29ee96d80a1cae88d878d8586b54e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records MiniMaxM2ForCausalLM, minimax_m2, dtype bfloat16, hidden size 3072, intermediate size 1536, 62 layers, 48 attention heads, 8 KV heads, head_dim 128, 256 local experts, 8 experts per token, no shared expert, max_position_embeddings 196608, sliding_window null, rope_theta 5000000, sigmoid routing with routing bias, use_qk_norm true, use_mtp true, and ModelOpt quantization metadata." }, { "label": "MiniMax M2.7 NVFP4 hf_quant_config", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4/raw/db821d7a3ce29ee96d80a1cae88d878d8586b54e/hf_quant_config.json", "source_type": "config", "supports": [ "serving", "weight_format", "quantized_module_scope" ], "notes": "hf_quant_config records quant_algo NVFP4, quant_method modelopt, 4-bit float weights and input activations with group_size 16, and ignore patterns for lm_head, every model.layers.*.block_sparse_moe.gate, and every model.layers.*.self_attn* module." }, { "label": "MiniMax M2.7 FP8 base profile/config comparison", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/d494266a4affc0d2995ba1fa35c8481cbd84294b/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison against the audited MiniMaxAI/MiniMax-M2.7 profile found matching architecture, model type, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, expert count, experts per token, rope theta, QK norm, and full-context sliding_window behavior. The NVFP4 repo sets max_position_embeddings 196608 while the FP8 base config sets 204800." }, { "label": "MiniMax M2.7 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4/resolve/db821d7a3ce29ee96d80a1cae88d878d8586b54e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The resolved safetensors index maps 191069 tensors across 25 model shards plus model-inputscales.safetensors. Range-read headers found 134.401725368 GB of tensor payload and 0.024089864 GB of safetensors header bytes. Payload dtypes are U8 112.340238336 GB, F8_E4M3 14.042529792 GB, BF16 8.018576384 GB, and F32 0.000380856 GB. model.embed_tokens.weight is BF16 [200064, 3072] and contributes 1.229193216 GB resident-only. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Non-expert layer tensors, model.norm.weight, and lm_head.weight total 6.789383168 GB. Routed expert tensors plus indexed F32 input_scale tensors total 126.383148984 GB, averaging 0.49368417571875 GB per expert across 256 experts." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, hf_quant_config, audited MiniMax M2.7 base profile comparison, safetensors index, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "This profile models ordinary text decode for the NVFP4 package. It deliberately includes the indexed input-scale sidecar as routed expert traffic and does not include unrelated amax/calibration backup safetensors that are present in the repo but not referenced by model.safetensors.index.json." }, { "id": "lyf--qwen3-6-35b-a3b-uncensored-hauhaucs-aggressive-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4", "title": "lyf Qwen3.6 35B A3B Uncensored HauhauCS Aggressive NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for the lyf conservative NVFP4 package of Qwen3.6 35B A3B Uncensored HauhauCS Aggressive.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "relation": "quantized", "source": "Hugging Face base_model metadata, served config comparison, model card, quantization recipe, official Qwen base config comparison, and direct safetensors header review", "config_compatible": true, "notes": "The lyf artifact records HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive as its quantized base. That base is a GGUF package derived from Qwen/Qwen3.6-35B-A3B. Because the HauhauCS base is GGUF-only in this audit environment, architecture compatibility is anchored on the lyf served config and the official Qwen/Qwen3.6-35B-A3B config; manual comparison found matching checked text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 23.336710112, "main_resident_weight_gb": 21.426448896, "auxiliary_resident_weight_gb": 1.910261216, "fixed_weight_gb": 3.306809856, "routed_expert_weight_gb": 0.07077984, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "single_safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through language_model and lm_head, excluding resident-only input embedding and visual tensors", "auxiliary_scope": "visual tensors and language_model.model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Direct range-read of the single safetensors header found 124306 tensors totaling 23.336710112 GB. Ordinary text resident tensors, defined as language_model excluding embed_tokens plus lm_head, total 21.426448896 GB. Auxiliary resident tensors, defined as visual plus language_model.model.embed_tokens.weight, total 1.910261216 GB. Main routed expert tensors total 18.119639040 GB and divide exactly into 256 byte-uniform expert indexes of 0.070779840 GB. Fixed ordinary text traffic totals 3.306809856 GB. The package has no resident MTP tensors despite the card noting that MTP tensors were copied from Qwen before quantization." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with full_attention_interval 4, giving 10 full-context attention layers. The model card's vLLM commands explicitly use FP8 KV cache, so this component uses the profile serving FP8 byte width." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The lyf card says linear_attn remains BF16 for best quality, so quantizing other weights and using FP8 full-attention KV does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower. This profile models ordinary cached text decode through the language model and output head, with resident-only multimodal tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.664727181660742, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-llm-compressor-nvfp4-fp8-kv-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored NVFP4 packed weights, FP8 scale tensors, BF16 unquantized tensors, F32 scalar scales, FP8 full-attention KV, and BF16/F32 DeltaNet state bytes. Activation quantization, dequantization, compute overhead, write traffic, image/video prefill, and router/expert compute are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights and activations with kv_cache_scheme null. The model card's vLLM examples use --kv-cache-dtype fp8, so this profile charges FP8 full-attention KV cache bytes." }, "evidence": [ { "label": "lyf Qwen3.6 35B HauhauCS NVFP4 API metadata", "url": "https://huggingface.co/api/models/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "serving" ], "notes": "At commit c9eb0a997ecfc688c468146404606ef6df49555f, the current API records a public non-gated Apache-2.0 image-text-to-text repo with qwen3_5_moe, qwen3.6, NVFP4, compressed-tensors, quantized, vLLM, MoE, multimodal, Blackwell, RTX 5090, uncensored, endpoints_compatible, 8-bit, and region:us tags. Current downloads are 136193. The API safetensors block records F32 61760, BF16 2496468336, F8_E4M3 2038169600, U8 16305356800, and total 20840056496 storage-accounting entries." }, { "label": "lyf Qwen3.6 35B HauhauCS NVFP4 model card", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/raw/c9eb0a997ecfc688c468146404606ef6df49555f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "kv_store_format", "runtime_format", "quantization_scope" ], "notes": "The card identifies the base as HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive, original model Qwen/Qwen3.6-35B-A3B, architecture as Qwen3.5 MoE with 35B total and 3B active parameters, 256 experts with 8 routed plus 1 shared expert, and conservative NVFP4 W4A4 quantization with linear_attn and MTP kept in BF16. It documents vLLM serving with compressed-tensors, NVFP4 Marlin GEMM, and --kv-cache-dtype fp8." }, { "label": "lyf Qwen3.6 35B HauhauCS NVFP4 config", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/raw/c9eb0a997ecfc688c468146404606ef6df49555f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "quantization_ignore_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, tie_word_embeddings false, a 27-layer vision_config, and kv_cache_scheme null. The compressed-tensors config targets Linear with nvfp4-pack-quantized format and a 341-entry ignore list covering visual modules, linear_attn projections, mlp.gate, shared_expert_gate, and MTP modules." }, { "label": "lyf Qwen3.6 35B HauhauCS NVFP4 quantization recipe", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/raw/c9eb0a997ecfc688c468146404606ef6df49555f/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantization_ignore_scope" ], "notes": "The recipe targets Linear with scheme NVFP4 and ignores lm_head, visual modules, mlp.gate, mlp.shared_expert_gate, linear_attn modules, and MTP modules." }, { "label": "HauhauCS Qwen3.6 35B A3B Uncensored Aggressive API metadata", "url": "https://huggingface.co/api/models/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "source_type": "model_card", "supports": [ "base_model_proof", "lineage" ], "notes": "The HauhauCS base repo is a public Apache-2.0 GGUF package derived from Qwen/Qwen3.6-35B-A3B and currently records GGUF architecture qwen35moe, context_length 262144, and region:us metadata. It does not provide the served safetensors config needed for this NVFP4 package, so the lyf served config and official Qwen config anchor the architecture audit." }, { "label": "Qwen3.6 35B A3B official base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison against the official base config found matching checked text_config and vision_config geometry fields. The lyf artifact adds compressed-tensors NVFP4 quantization metadata while preserving the base architecture." }, { "label": "lyf Qwen3.6 35B HauhauCS NVFP4 safetensors header", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/resolve/c9eb0a997ecfc688c468146404606ef6df49555f/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Direct range-read of the single safetensors header found 124306 tensors totaling 23.336710112 GB, with linked file size 23.354242416 GB and 0.017532304 GB of header/container overhead outside tensor payloads. Stored bytes split into F32 0.000247040 GB, BF16 4.992936672 GB, F8_E4M3 2.038169600 GB, and U8 16.305356800 GB. Ordinary text resident tensors, defined as language_model excluding embed_tokens plus lm_head, sum to 21.426448896 GB. Auxiliary resident tensors, defined as visual plus language_model.model.embed_tokens.weight, sum to 1.910261216 GB. Routed expert tensors sum to 18.119639040 GB and divide exactly into 256 uniform expert indexes of 0.070779840 GB. Fixed ordinary text traffic sums to 3.306809856 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caches conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to the value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current Hugging Face API metadata, pinned model card, served NVFP4 config, quantization recipe, HauhauCS GGUF base metadata, official Qwen base config, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile intentionally models ordinary text decode for the lyf conservative NVFP4 body only. It does not model image/video prefill, the GGUF source package, or any future speculative decode path." }, { "id": "maziyarpanahi--deepseek-r1-0528-qwen3-8b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF", "title": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the row-selected Q2_K GGUF artifact of DeepSeek R1 0528 Qwen3 8B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, stale repo-config check, and audited BF16 DeepSeek base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of deepseek-ai/DeepSeek-R1-0528-Qwen3-8B. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q2_K GGUF header records the same Qwen3ForCausalLM geometry as the audited BF16 base profile: 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 key/value head dimension, 131072 context, YARN RoPE scaling from 32768 to 131072, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "deepseek-r1-0528-qwen3-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 3.281731712, "swept_weight_gb": 3.071574016, "auxiliary_resident_weight_gb": 0.210157696, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for DeepSeek-R1-0528-Qwen3-8B.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 3.281731712 GB. GGUF tensor spans total 3.275776000 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005955712 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 36 Qwen3 decoder layers, 8 KV heads, 128-dimensional key/value heads, and 131072 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the row-selected Q2_K GGUF artifact. FP16 and other quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.40066386810963944, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets DeepSeek-R1-0528-Qwen3-8B.Q2_K.gguf because the catalog row is the 2-bit GGUF entry. The live HF API gguf.totalFileSize currently matches the FP16 sibling instead, so exact selected-file header bytes are authoritative. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 3d7c2a7c46f7cd566b7318e49986dda9a6d7d190, the API records a public non-gated GGUF text-generation repo with base_model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, base_model:quantized metadata, quantized tags, region:us, 124510 downloads, GGUF architecture qwen3, context_length 131072, gguf.total 8190735360, and gguf.totalFileSize 16388042880. The API totalFileSize matches the fp16 sibling, while this profile targets the Q2_K file for the catalog's 2-bit row." }, { "label": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base-model profile record MIT licensing." }, { "label": "DeepSeek R1 0528 Qwen3 8B audited BF16 base profile", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The audited BF16 base profile records Qwen3ForCausalLM, bfloat16, MIT licensing, 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 key/value head dimension, 131072 max positions, no sliding window, and untied embeddings with separate model.embed_tokens.weight and lm_head.weight." }, { "label": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/tree/3d7c2a7c46f7cd566b7318e49986dda9a6d7d190", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found Q2_K 3.281731712 GB, Q3_K_M 4.124160128 GB, Q3_K_L 4.431392896 GB, Q4_K_M 5.027782784 GB, Q5_K_M 5.851111552 GB, Q6_K 6.725898368 GB, and fp16 16.388042880 GB. The selected Q2_K artifact is the standard 2-bit GGUF file for this row." }, { "label": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/resolve/3d7c2a7c46f7cd566b7318e49986dda9a6d7d190/DeepSeek-R1-0528-Qwen3-8B.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 32 metadata entries and 399 tensors. The linked file is 3.281731712 GB. Tensor spans sum to 3.275776000 GB: output.weight 0.510504960 GB, token_embd.weight 0.204201984 GB, blk.* tensors 2.561052672 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.005955712 GB. Stored tensor spans split into Q2_K 1.641013248 GB, Q3_K 1.038090240 GB, Q4_K 0.084934656 GB, Q6_K 0.510504960 GB, and F32 0.001232896 GB. The header records general.architecture qwen3, MIT license, qwen3.block_count 36, context_length 131072, embedding_length 4096, feed_forward_length 12288, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000, YARN scaling factor 4 from 32768 original context, and a separate output.weight tensor." }, { "label": "MaziyarPanahi DeepSeek R1 0528 Qwen3 8B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/raw/3d7c2a7c46f7cd566b7318e49986dda9a6d7d190/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base profile and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the row-selected DeepSeek R1 0528 Qwen3 8B Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--deepseek-v3-0324-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/DeepSeek-V3-0324-GGUF", "title": "MaziyarPanahi DeepSeek V3 0324 GGUF IQ1_M Split", "summary": "Audited memory-side text-decode bounds profile for the API-selected 9-part IQ1_M split GGUF artifact of DeepSeek V3 0324.", "model_family": "deepseek-v3-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V3-0324", "relation": "quantized", "source": "Hugging Face model card/API metadata, DeepSeek V3 0324 config, official local inference adapter review, selected split-GGUF header metadata, and selected linked-object size checks", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a quantized derivative of deepseek-ai/DeepSeek-V3-0324. The selected GGUF headers match the audited DeepSeek V3 0324 layer count, MLA ranks, expert counts, routing shape, and context length." }, "architecture": { "canonical_architecture_id": "deepseek-v3-0324", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 148.882857728, "main_resident_weight_gb": 148.573487104, "auxiliary_resident_weight_gb": 0.309370624, "fixed_weight_gb": 4.708859904, "routed_expert_weight_gb": 0.5619712, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected 9-part IQ1_M split GGUF linked-file size and API GGUF logical tensor parameters", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, attention tensors, shared/dense FFN tensors, router tensors, and expected distinct routed expert tensor spans from the selected IQ1_M split; token_embd.weight is resident-only input embedding traffic", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, headers, and file overhead are resident in the selected split artifact but not swept for ordinary text decode", "shared_expert_notes": "The selected GGUF header records expert_shared_count 1. Dense leading-block FFN tensors and shared-expert tensors are charged in fixed_weight_gb because they are always-on traffic.", "notes": "HF API gguf.total is 671.026419200B parameters and gguf.totalFileSize selects the 9-part DeepSeek-V3-0324.IQ1_M split. Range-reads of all nine GGUF v3 shard headers found 1025 tensors. Linked split files total 148.882857728 GB. Tensor spans total 148.877553664 GB, while metadata/header/tokenizer/alignment overhead accounts for 0.005304064 GB. Routed expert tensors total 143.864627200 GB across layers 3-60 and 256 expert indexes, or 0.561971200 GB per expert index." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Official DeepSeek-V3 absorb-MLA cache coefficient: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 BF16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same MLA latent plus RoPE cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "The selected GGUF header records the DeepSeek2 MLA geometry. As with the audited official DeepSeek V3 0324 profile, this profile follows the official absorb MLA local-serving path rather than a generic expanded K/V cache." }, "notes": "This profile models ordinary text decode for the API-selected IQ1_M split GGUF artifact. It does not substitute the IQ1_S, Q2_K, or Q3_K_S split families." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.22187331745521832, "kv_store_format": "mla_bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "mla_bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq1-m-deepseek-v3-absorb-mla-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected split-GGUF tensor spans for weight traffic and selected linked-file bytes for residency. GGUF split loading overhead, kernels, dequantization, scheduler behavior, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The API-selected artifact is the 9-part IQ1_M split because the split linked-file sum exactly matches HF API gguf.totalFileSize. The KV/state adapter follows the official DeepSeek V3 absorb MLA serving path already used for the audited official FP8 profile." }, "evidence": [ { "label": "MaziyarPanahi DeepSeek V3 0324 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/DeepSeek-V3-0324-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit d77054856dd5bb226f066fe76063a3691ffc3cfd, the live API records a public non-gated GGUF repo with base_model deepseek-ai/DeepSeek-V3-0324, base_model_relation quantized, MIT license metadata, region:us, imatrix metadata, 108352 downloads, GGUF architecture deepseek2, context length 163840, gguf.total 671026419200, and gguf.totalFileSize 148882857728." }, { "label": "MaziyarPanahi DeepSeek V3 0324 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-V3-0324-GGUF/raw/d77054856dd5bb226f066fe76063a3691ffc3cfd/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The card metadata records base_model deepseek-ai/DeepSeek-V3-0324, quantized_by MaziyarPanahi, GGUF/llama.cpp packaging, MIT licensing, and available IQ1_M, IQ1_S, Q2_K, and Q3_K_S split artifacts." }, { "label": "DeepSeek V3 0324 audited base profile evidence", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/raw/e9b33add76883f293d6bf61f6bd89b497e80e335/config.json", "source_type": "config", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "compressed_state", "max_context_tokens" ], "notes": "The audited official profile records the matching DeepSeek V3 0324 config: 61 hidden layers, 3 initial dense layers, 256 routed experts, 8 experts per token, 1 shared expert, hidden size 7168, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, and 163840 max position embeddings. Official local inference uses the absorb MLA cache path charged here." }, { "label": "MaziyarPanahi DeepSeek V3 0324 GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-V3-0324-GGUF/tree/d77054856dd5bb226f066fe76063a3691ffc3cfd", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Exact HEAD checks found the nine IQ1_M split shards total 148.882857728 GB, exactly matching API gguf.totalFileSize. The alternate split families total 133.555674368 GB for IQ1_S, 244.028344832 GB for Q2_K, and 289.082589184 GB for Q3_K_S; they are not selected by this profile." }, { "label": "MaziyarPanahi DeepSeek V3 0324 IQ1_M split GGUF header audit", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-V3-0324-GGUF/resolve/d77054856dd5bb226f066fe76063a3691ffc3cfd/DeepSeek-V3-0324.IQ1_M.gguf-00001-of-00009.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_adapter", "weight_format" ], "notes": "Range-reads of the selected IQ1_M split GGUF v3 headers found split.count 9 and split.tensors.count 1025. Header tensor spans across all nine shards sum to 148.877553664 GB: routed expert tensors 143.864627200 GB, attention tensors 2.778773504 GB, router/scale/norm FFN tensors 0.989442048 GB, output.weight 0.637091840 GB, token_embd.weight 0.304066560 GB, shared/dense FFN tensors 0.303464448 GB, blk other tensors 0.000059392 GB, and output_norm.weight 0.000028672 GB. Metadata/header/tokenizer/alignment overhead across the split accounts for 0.005304064 GB. Stored tensor spans split into IQ1_M 143.053711360 GB, Q2_K 2.910007296 GB, IQ2_XXS 1.846935552 GB, Q5_K 0.637091840 GB, and F32 0.429807616 GB. The first shard header records deepseek2.block_count 61, context_length 163840, embedding_length 7168, feed_forward_length 18432, attention.head_count 128, attention.head_count_kv 128, q_lora_rank 1536, kv_lora_rank 512, key_length 192, value_length 128, expert_count 256, expert_used_count 8, expert_shared_count 1, leading_dense_block_count 3, rope.dimension_count 64, and yarn scaling factor 40." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, HF CLI model info, model card metadata, the audited official DeepSeek V3 0324 serving adapter, linked-object HEAD checks, and direct GGUF header/tensor-index range reads of all nine selected IQ1_M split shards." }, "notes": "Use this profile for the API-selected IQ1_M split GGUF artifact. Do not infer the IQ1_S, Q2_K, or Q3_K_S split footprints unless the workload profile explicitly selects and audits those split families." }, { "id": "maziyarpanahi--firefunction-v2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/firefunction-v2-GGUF", "title": "MaziyarPanahi Firefunction V2 GGUF IQ1_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected IQ1_M GGUF artifact of Firefunction V2.", "model_family": "llama3-70b-firefunction-v2-gguf", "base_model_proof": { "base_model": "fireworks-ai/llama-3-firefunction-v2", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected IQ1_M GGUF header metadata, package-config absence check, and pinned Fireworks base config", "config_compatible": true, "notes": "The repo tags and existing catalog row identify this package as a GGUF derivative of fireworks-ai/llama-3-firefunction-v2; the package cardData also mentions fireworks-ai/firefunction-v2. The canonical llama-3-firefunction-v2 base repo is public and records LlamaForCausalLM with the same checked geometry as the selected GGUF header: 80 layers, hidden size 8192, intermediate size 28672, 64 attention heads, 8 KV heads, 8192 max positions, and untied embeddings." }, "architecture": { "canonical_architecture_id": "llama3-70b-firefunction-v2", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 16.751198848, "swept_weight_gb": 16.39858176, "auxiliary_resident_weight_gb": 0.352617088, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for firefunction-v2.IQ1_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected IQ1_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, imatrix metadata, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected IQ1_M linked file is 16.751198848 GB. GGUF tensor spans total 16.743333888 GB, while GGUF metadata, tokenizer, header, imatrix metadata, and file alignment account for 0.007864960 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes IQ1_M, IQ2_XXS, Q2_K, Q4_K, Q5_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header and base config record 80 decoder blocks, 8 KV heads, 128-dimensional key/value heads, 8192 context, and no sliding-window cap. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected IQ1_M GGUF artifact." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.23742478857506516, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq1-m-imatrix-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, kernels, scheduler behavior, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets firefunction-v2.IQ1_M.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Firefunction V2 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/firefunction-v2-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8, the API records a public non-gated Transformers/GGUF text-generation repo with base_model tags for fireworks-ai/llama-3-firefunction-v2, cardData base_model fireworks-ai/firefunction-v2, license llama3, imatrix, region:us, 111313 downloads, GGUF architecture llama, context_length 8192, gguf.total 70553706496, and gguf.totalFileSize 16751198848. The API totalFileSize matches firefunction-v2.IQ1_M.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Firefunction V2 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF/raw/e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records the Fireworks base model relationship, quantized_by MaziyarPanahi, GGUF/llama.cpp runtime guidance, function-calling tags, and imatrix quantization metadata." }, { "label": "Fireworks Llama 3 Firefunction V2 base API metadata", "url": "https://huggingface.co/api/models/fireworks-ai/llama-3-firefunction-v2", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit 7954e120254ced10ac26a2dd346679d7328fa9ae, the base repo is public, non-gated, Transformers text-generation tagged, region:us tagged, license llama3, and records F16 safetensors total 70553706496 parameters." }, { "label": "Fireworks Llama 3 Firefunction V2 base config", "url": "https://huggingface.co/fireworks-ai/llama-3-firefunction-v2/raw/7954e120254ced10ac26a2dd346679d7328fa9ae/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The base config records LlamaForCausalLM, float16 dtype, 80 layers, hidden size 8192, intermediate size 28672, 64 attention heads, 8 KV heads, 8192 max positions, rope_theta 500000, use_cache true, tie_word_embeddings false, and vocab size 128256." }, { "label": "MaziyarPanahi Firefunction GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF/tree/e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF linked-object metadata found IQ1_M 16.751198848 GB, IQ1_S 15.343485568 GB, IQ2_XS 21.142110848 GB, IQ3_XS 29.307732608 GB, IQ4_XS 37.902664320 GB, Q2_K 26.375111296 GB, Q3_K_S 30.912053888 GB, Q3_K_M 34.267497088 GB, Q3_K_L 37.140595328 GB, Q4_K_S 40.347222656 GB, Q4_K_M 42.520396416 GB, Q5_K_S 48.657449600 GB, and Q5_K_M 49.949819520 GB. The selected IQ1_M artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Firefunction IQ1_M GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF/resolve/e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8/firefunction-v2.IQ1_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 28 metadata entries and 723 tensors. The linked file is 16.751198848 GB, with tensor data beginning at byte 7864960. Tensor spans sum to 16.743333888 GB: output.weight 0.722337792 GB, output_norm.weight 0.000032768 GB, token_embd.weight 0.344752128 GB, and blk.* tensors 15.676211200 GB. Metadata/tokenizer/header/file overhead accounts for 0.007864960 GB. Stored tensor bytes split into IQ1_M 13.138657280 GB, IQ2_XXS 1.384120320 GB, Q2_K 1.115455488 GB, Q4_K 0.377487360 GB, Q5_K 0.722337792 GB, and F32 0.005275648 GB. The header records general.architecture llama, 80 blocks, 8192 context, 8192 embedding length, 28672 feed-forward length, 64 attention heads, 8 KV heads, 128 RoPE/head dimension, rope.freq_base 500000, vocab size 128256, and separate output.weight." }, { "label": "MaziyarPanahi Firefunction GGUF package config absence", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF/raw/e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package has no repo-local config.json; the raw config request returned HTTP 404. This profile therefore uses the selected GGUF header plus the pinned Fireworks base config as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, linked-file size metadata for GGUF siblings, package config absence check, and a direct GGUF header/tensor-index range read of the selected IQ1_M artifact." }, "notes": "Use this profile for the API-selected Firefunction V2 IQ1_M GGUF artifact. Do not infer footprints for the other GGUF sibling quantizations or split FP16/Q6/Q8 files from this profile." }, { "id": "maziyarpanahi--gemma-2-2b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/gemma-2-2b-it-GGUF", "title": "MaziyarPanahi Gemma 2 2B IT GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Gemma 2 2B IT.", "model_family": "gemma2-dense-gguf", "base_model_proof": { "base_model": "google/gemma-2-2b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, existing gated Google base-profile checks, selected GGUF header metadata, and the Transformers Gemma 2 implementation", "config_compatible": false, "notes": "The GGUF repo metadata identifies google/gemma-2-2b-it as the quantized base. The repo-local config.json is a stale 31-byte Mistral stub, and the Google base config is gated in this audit environment, so this profile uses the selected public GGUF header and pinned Gemma 2 implementation defaults as direct architecture evidence." }, "architecture": { "canonical_architecture_id": "gemma-2-2b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.614341888, "swept_params_b": 2.614341888, "auxiliary_resident_params_b": 0, "resident_weight_gb": 5.235213952, "swept_weight_gb": 5.229167616, "auxiliary_resident_weight_gb": 0.006046336, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-2-2b-it.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected FP16 linked file is 5.235213952 GB. Header tensor spans total 5.229167616 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.006046336 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 13, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The pinned Transformers Gemma 2 config alternates sliding_attention on zero-indexed even layers and full_attention on zero-indexed odd layers. With 26 layers, this gives 13 full-context attention layers." }, { "kind": "sliding_window", "layers": 13, "kv_heads": 4, "head_dim": 256, "window_tokens": 4096, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records gemma2.attention.sliding_window 4096, and the pinned Gemma 2 attention path applies that window only to sliding_attention layers." } ], "notes": "Hybrid Gemma 2 text attention is represented as explicit full-context plus sliding-window KV components. The selected GGUF artifact does not declare quantized KV cache, so Bounds Engine v1 charges llama.cpp-style FP16 K/V cache storage and read traffic." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. It uses public GGUF metadata and pinned Transformers Gemma 2 implementation evidence because the Google base config is gated and the repo-local config stub is stale." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0024978278586953, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-fp16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many smaller GGUF quantizations. This profile intentionally targets gemma-2-2b-it.fp16.gguf because the HF API gguf.totalFileSize exactly matches that linked object, following the selected-artifact rule used by the existing GGUF profiles." }, "evidence": [ { "label": "MaziyarPanahi Gemma 2 2B IT GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-2-2b-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit bd652eddf75b473fe86639b6b927e06972277d1a, the API records a public non-gated GGUF repo with base_model google/gemma-2-2b-it, region:us, 119685 downloads, GGUF architecture gemma2, 8192 context length, gguf.total 2614341888, and gguf.totalFileSize 5235213952. The API totalFileSize exactly matches gemma-2-2b-it.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Gemma 2 2B IT GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/raw/bd652eddf75b473fe86639b6b927e06972277d1a/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "quantization_recipe" ], "notes": "The card records this as a GGUF package by Maziyar Panahi for google/gemma-2-2b-it, with 2-bit through 8-bit tags. It does not override the HF API-selected FP16 artifact." }, { "label": "MaziyarPanahi Gemma 2 2B IT repo config stub", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/raw/bd652eddf75b473fe86639b6b927e06972277d1a/config.json", "source_type": "manual_review", "supports": [ "stale_config" ], "notes": "The repo-local config.json is a 31-byte stub declaring model_type mistral. The selected GGUF header and API gguf block identify the model as gemma2, so the stub is not used for bounds geometry." }, { "label": "Google Gemma 2 2B IT gated base profile", "url": "https://huggingface.co/google/gemma-2-2b-it", "source_type": "manual_review", "supports": [ "base_model_proof", "unsupported_gated_base" ], "notes": "The existing local unsupported profile for google/gemma-2-2b-it records API safetensors total 2614341888 parameters at commit 299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8, but config and tensor headers are gated. This public GGUF derivative is audited from its selected artifact instead of inferring from the gated base profile." }, { "label": "MaziyarPanahi Gemma 2 2B IT GGUF linked-object metadata", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/tree/bd652eddf75b473fe86639b6b927e06972277d1a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata records gemma-2-2b-it.fp16.gguf 5.235213952 GB, Q8_0 2.784495456 GB, Q6_K 2.151393120 GB, Q5_K_M 1.923278688 GB, Q5_K_S 1.882543968 GB, Q4_K_M 1.708582752 GB, Q4_K_S 1.638651744 GB, IQ4_XS 1.566250848 GB, Q3_K_L 1.550436192 GB, Q3_K_M 1.461667680 GB, Q3_K_S 1.360660320 GB, IQ3_XS 1.314211680 GB, Q2_K 1.229829984 GB, IQ2_XS 1.002544992 GB, IQ1_M 0.873797472 GB, IQ1_S 0.832159584 GB, and imatrix.dat 0.002375548 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Gemma 2 2B IT FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/resolve/bd652eddf75b473fe86639b6b927e06972277d1a/gemma-2-2b-it.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 288 tensors. The linked file is 5.235213952 GB. Tensor spans sum to 5.229167616 GB: token_embd.weight 1.179648000 GB, blk.* tensors 4.049510400 GB, and output_norm.weight 0.000009216 GB. Metadata/tokenizer/header/file overhead accounts for 0.006046336 GB. Stored tensor bytes split into 5.228199936 GB F16 and 0.000967680 GB F32. The header records gemma2.block_count 26, context_length 8192, embedding_length 2304, feed_forward_length 9216, attention.head_count 8, attention.head_count_kv 4, attention key/value length 256, attention.sliding_window 4096, and no separate output.weight tensor." }, { "label": "Transformers Gemma 2 config and attention implementation", "url": "https://github.com/huggingface/transformers/blob/b70d02fc724d04c916832ca4ead03ff05e8fb1ee/src/transformers/models/gemma2/configuration_gemma2.py", "source_type": "config", "supports": [ "layer_pattern", "kv_adapter", "sliding_window" ], "notes": "The pinned Gemma2Config defaults match the GGUF geometry: 26 layers, 8 attention heads, 4 KV heads, 256 head dimension, 8192 max positions, and 4096 sliding window. Its __post_init__ creates sliding_attention for zero-indexed even layers and full_attention for zero-indexed odd layers. The pinned Gemma2Attention implementation sets self.sliding_window only for sliding_attention layers and passes that window to the attention kernel." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, stale config-stub check, existing gated-base profile evidence, expanded linked-file metadata, a direct GGUF header/tensor-index range read of the selected FP16 artifact, and pinned Transformers Gemma 2 config/attention implementation review." }, "notes": "Use this profile for the API-selected Gemma 2 2B IT FP16 GGUF artifact. Do not silently substitute the smaller Q4_K_M or other quantized artifacts; those require separate profiles with their own selected artifact bytes." }, { "id": "maziyarpanahi--gemma-3-12b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/gemma-3-12b-it-GGUF", "title": "MaziyarPanahi Gemma 3 12B IT GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Gemma 3 12B IT.", "model_family": "gemma3-dense-gguf", "base_model_proof": { "base_model": "google/gemma-3-12b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, package config conflict check, gated base access check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-12b-it. The Google base repo remains gated in this audit environment, and the package config.json incorrectly records only model_type mistral, so this profile uses the selected public GGUF header as the architecture source instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-12b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.765788416, "swept_params_b": 11.765788416, "resident_weight_gb": 23.539658528, "swept_weight_gb": 23.533108224, "auxiliary_resident_weight_gb": 0.006550304, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-12b-it.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans; token_embd.weight is also the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The selected FP16 linked file is 23.539658528 GB. GGUF tensor spans total 23.533108224 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006550304 GB. The absence of output.weight means token_embd.weight remains in swept decode traffic as the tied output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 8, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 48 blocks, that yields eight full-context global layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 1024 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head and window metadata. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly selects a quantized KV cache." }, "notes": "Dense Gemma3 GGUF profile audited from the selected FP16 artifact, not from the gated Google raw config. The selected artifact contains text tensors only: token embedding, 48 blk.* layers, and output norm." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, image prefill, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches gemma-3-12b-it.fp16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant FP16 tensor format." }, "evidence": [ { "label": "MaziyarPanahi Gemma 3 12B IT GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-12b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 0e7c92b7b0890410acb9638a0107c8722a6ab1df, the live API records a public non-gated GGUF text-generation repo with base_model google/gemma-3-12b-it, base_model:quantized metadata, region:us, 117347 downloads, GGUF architecture gemma3, context_length 131072, gguf.total 11765788416, and gguf.totalFileSize 23539658528. The API totalFileSize matches gemma-3-12b-it.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Gemma 3 12B IT GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/raw/0e7c92b7b0890410acb9638a0107c8722a6ab1df/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model google/gemma-3-12b-it, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. The card advertises multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a smaller default." }, { "label": "MaziyarPanahi Gemma 3 12B IT GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/raw/0e7c92b7b0890410acb9638a0107c8722a6ab1df/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral, which conflicts with the live API GGUF architecture and the selected GGUF header. This profile therefore does not use that config as architecture evidence." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 48 blocks recorded by the selected GGUF header." }, { "label": "Google Gemma 3 12B IT gated base config access check", "url": "https://huggingface.co/google/gemma-3-12b-it/raw/96b6f1eccf38110c56df3a15bffe176da04bfd80/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The raw base config returned HTTP 401 with an access-restricted response in this audit environment. The existing google/gemma-3-12b-it profile is therefore still an unsupported gated-base stub, and this GGUF profile does not claim a direct base-config comparison." }, { "label": "MaziyarPanahi Gemma 3 12B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/tree/0e7c92b7b0890410acb9638a0107c8722a6ab1df", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF CLI dry-run and exact HEAD checks found FP16 23.539658528 GB, Q8_0 12.509949728 GB, Q6_K 9.660608288 GB, Q5_K_M 8.444833568 GB, Q5_K_S 8.231759648 GB, Q4_K_M 7.300575008 GB, Q4_K_S 6.935129888 GB, Q3_K_L 6.479982368 GB, Q3_K_M 6.008614688 GB, Q3_K_S 5.458112288 GB, and Q2_K 4.768018208 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Gemma 3 12B IT FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/resolve/0e7c92b7b0890410acb9638a0107c8722a6ab1df/gemma-3-12b-it.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 40 metadata entries and 626 tensors. The linked file is 23.539658528 GB. Tensor spans sum to 23.533108224 GB: token_embd.weight 2.013265920 GB, blk.* tensors 21.519826944 GB, and output_norm.weight 0.000015360 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006550304 GB. Stored tensor spans split into F16 23.530045440 GB and F32 0.003062784 GB. The header records general.architecture gemma3, gemma3.block_count 48, context_length 131072, embedding_length 3840, feed_forward_length 15360, attention.head_count 16, attention.head_count_kv 8, attention key/value length 256, attention.sliding_window 1024, rope scaling factor 8, vocab_size 262144, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, HF CLI dry-run, model card metadata, package config conflict check, gated-base access check, selected linked-object HEAD checks, existing Gemma 3 layered-KV treatment, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Gemma 3 12B IT FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--gemma-3-1b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/gemma-3-1b-it-GGUF", "title": "MaziyarPanahi Gemma 3 1B IT GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Gemma 3 1B IT.", "model_family": "gemma3-dense-gguf", "base_model_proof": { "base_model": "google/gemma-3-1b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, Google base API metadata, package config conflict check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-1b-it. The Google base API is accessible for high-level metadata and confirms the Gemma license and 999885952 BF16 parameters, but the full raw base config remains gated in this audit environment. The package config.json incorrectly records only model_type mistral, so this profile uses the selected public GGUF header as the architecture source." }, "architecture": { "canonical_architecture_id": "gemma-3-1b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.999885952, "swept_params_b": 0.999885952, "resident_weight_gb": 2.0065736, "swept_weight_gb": 2.000040448, "auxiliary_resident_weight_gb": 0.006533152, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-1b-it.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans; token_embd.weight is also the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The selected FP16 linked file is 2.006573600 GB. GGUF tensor spans total 2.000040448 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006533152 GB. The absence of output.weight means token_embd.weight remains in swept decode traffic as the tied output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 4, "kv_heads": 1, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 26 blocks, that yields four full-context global layers." }, { "kind": "sliding_window", "layers": 22, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 512 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head and window metadata." }, "notes": "Dense Gemma3 GGUF profile audited from the selected FP16 artifact, not from the gated Google raw config." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches gemma-3-1b-it.fp16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant FP16 tensor format." }, "evidence": [ { "label": "MaziyarPanahi Gemma 3 1B IT GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-1b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 9aea24a60427bba55a903a1c4b526842d40fad3e, the live API records a public non-gated GGUF text-generation repo with base_model google/gemma-3-1b-it, base_model:quantized metadata, region:us, 122255 downloads, GGUF architecture gemma3, context_length 32768, gguf.total 999885952, and gguf.totalFileSize 2006573600. The API totalFileSize matches gemma-3-1b-it.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Gemma 3 1B IT GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/raw/9aea24a60427bba55a903a1c4b526842d40fad3e/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model google/gemma-3-1b-it, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. The card advertises multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a smaller default." }, { "label": "Google Gemma 3 1B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-1b-it", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "The live base API records a public but gated Gemma-license text-generation repo with base_model google/gemma-3-1b-pt, region:us, Gemma3ForCausalLM high-level config metadata, and safetensors BF16 total 999885952 parameters." }, { "label": "MaziyarPanahi Gemma 3 1B IT GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/raw/9aea24a60427bba55a903a1c4b526842d40fad3e/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral, which conflicts with the live API GGUF architecture and the selected GGUF header. This profile therefore does not use that config as architecture evidence." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 26 blocks recorded by the selected GGUF header." }, { "label": "MaziyarPanahi Gemma 3 1B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/tree/9aea24a60427bba55a903a1c4b526842d40fad3e", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 2.006573600 GB, Q8_0 1.069306400 GB, Q6_K 1.011738656 GB, Q5_K_M 0.851345696 GB, Q5_K_S 0.836399648 GB, Q4_K_M 0.806058272 GB, Q4_K_S 0.780993056 GB, Q3_K_L 0.751575584 GB, Q3_K_M 0.722416160 GB, Q3_K_S 0.688856096 GB, and Q2_K 0.689814560 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Gemma 3 1B IT FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/resolve/9aea24a60427bba55a903a1c4b526842d40fad3e/gemma-3-1b-it.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "max_context_tokens" ], "notes": "A 16MB range-read of the GGUF v3 header found 38 metadata entries and 340 tensors. The linked file is 2.006573600 GB. Tensor spans sum to 2.000040448 GB: token_embd.weight 0.603979776 GB, blk.* tensors 1.396056064 GB, and output_norm.weight 0.000004608 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006533152 GB. Stored tensor spans split into F16 1.999503360 GB and F32 0.000537088 GB. The header records general.architecture gemma3, gemma3.block_count 26, context_length 32768, embedding_length 1152, feed_forward_length 6912, attention.head_count 4, attention.head_count_kv 1, attention key/value length 256, attention.sliding_window 512, rope.freq_base 1000000, vocab_size 262144, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, Google base API metadata, package config conflict check, selected linked-object HEAD checks, existing Gemma 3 layered-KV treatment, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Gemma 3 1B IT FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--gemma-3-27b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/gemma-3-27b-it-GGUF", "title": "MaziyarPanahi Gemma 3 27B IT GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Gemma 3 27B IT.", "model_family": "gemma3-dense-gguf", "base_model_proof": { "base_model": "google/gemma-3-27b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, package config absence check, gated base access check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-27b-it. The Google base repo remains gated in this audit environment, and this GGUF package does not publish a config.json, so this profile uses the selected public GGUF header as the architecture source instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.00900224, "swept_params_b": 27.00900224, "resident_weight_gb": 10.503436704, "swept_weight_gb": 10.49687552, "auxiliary_resident_weight_gb": 0.006561184, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-27b-it.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected Q2_K artifact; token_embd.weight is charged as tied output-projection traffic because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The selected Q2_K linked file is 10.503436704 GB. GGUF tensor spans total 10.496875520 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006561184 GB. Tensor spans split into Q2_K, Q3_K, Q6_K, and F32 tensors. The header has token_embd.weight and output_norm.weight as the only non-block tensors and no separate output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 62 blocks, that yields 10 full-context global layers." }, { "kind": "sliding_window", "layers": 52, "kv_heads": 16, "head_dim": 128, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 1024 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head, head dimension, and window metadata. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly selects a quantized KV cache." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. The package also contains larger GGUF siblings that require separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3888865131213377, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-q3-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, scheduler behavior, dequantization, activation traffic, image prefill, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the Q2_K GGUF because gguf.totalFileSize exactly matches gemma-3-27b-it.Q2_K.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param only summarizes selected linked file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "MaziyarPanahi Gemma 3 27B IT GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-27b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit f1a115f273da292b54e6ef203f5c4f8645f87d5f, the live API records a public non-gated GGUF text-generation repo with base_model google/gemma-3-27b-it, base_model:quantized metadata, region:us, 114543 downloads, GGUF architecture gemma3, context_length 131072, gguf.total 27009002240, and gguf.totalFileSize 10503436704. The API totalFileSize matches gemma-3-27b-it.Q2_K.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Gemma 3 27B IT GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF/raw/f1a115f273da292b54e6ef203f5c4f8645f87d5f/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model google/gemma-3-27b-it, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. It advertises multiple lower-bit quantizations and does not override the API-selected Q2_K artifact with a larger default." }, { "label": "MaziyarPanahi Gemma 3 27B IT GGUF package config absence check", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF/raw/f1a115f273da292b54e6ef203f5c4f8645f87d5f/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package siblings list does not include config.json, and a raw config request returned HTTP 404. This profile therefore uses the selected GGUF header, not a package config, as architecture evidence." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 62 blocks recorded by the selected GGUF header." }, { "label": "Google Gemma 3 27B IT gated base config access check", "url": "https://huggingface.co/google/gemma-3-27b-it/raw/005ad3404e59d6023443cb575daa05336842228a/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The raw base config returned HTTP 401 with an access-restricted response in this audit environment. The existing google/gemma-3-27b-it profile is therefore still an unsupported gated-base stub, and this GGUF profile does not claim a direct base-config comparison." }, { "label": "MaziyarPanahi Gemma 3 27B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF/tree/f1a115f273da292b54e6ef203f5c4f8645f87d5f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Exact HEAD checks found Q2_K 10.503436704 GB, Q3_K_S 12.167330208 GB, Q3_K_M 13.437356448 GB, Q3_K_L 14.543178144 GB, Q4_K_S 15.673772448 GB, Q4_K_M 16.546404768 GB, Q5_K_S 18.766907808 GB, Q5_K_M 19.271391648 GB, Q6_K 22.166690208 GB, and Q8_0 28.707604896 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Gemma 3 27B IT Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF/resolve/f1a115f273da292b54e6ef203f5c4f8645f87d5f/gemma-3-27b-it.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 40 metadata entries and 808 tensors. The linked file is 10.503436704 GB. Tensor spans sum to 10.496875520 GB: token_embd.weight 1.156055040 GB, blk.* tensors 9.340798976 GB, and output_norm.weight 0.000021504 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006561184 GB. Stored tensor spans split into Q2_K 5.375655936 GB, Q3_K 3.959746560 GB, Q6_K 1.156055040 GB, and F32 0.005417984 GB. The header records general.architecture gemma3, gemma3.block_count 62, context_length 131072, embedding_length 5376, feed_forward_length 21504, attention.head_count 32, attention.head_count_kv 16, attention key/value length 128, attention.sliding_window 1024, rope scaling factor 8, Gemma license metadata, vocab_size 262144, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, package config absence check, gated-base access check, linked-object HEAD checks, existing Gemma 3 layered-KV treatment, and a direct GGUF header/tensor-index range read of the selected Q2_K artifact." }, "notes": "Use this profile for the API-selected Gemma 3 27B IT Q2_K GGUF artifact. Do not infer the larger sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--gemma-3-4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/gemma-3-4b-it-GGUF", "title": "MaziyarPanahi Gemma 3 4B IT GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Gemma 3 4B IT.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-4b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata and the public selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-4b-it. That base repo remains gated in this audit environment, and the package config.json incorrectly records only model_type mistral, so this profile uses the selected public GGUF header as the architecture source instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.880099328, "swept_params_b": 3.880099328, "resident_weight_gb": 7.767474368, "swept_weight_gb": 7.760934912, "auxiliary_resident_weight_gb": 0.006539456, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-4b-it.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected FP16 artifact; token_embd.weight is charged as tied output-projection traffic because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The selected FP16 linked file is 7.767474368 GB. GGUF tensor spans total 7.760934912 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006539456 GB. Tensor spans split into 7.759462400 GB F16 tensors and 0.001472512 GB F32 tensors. The header has token_embd.weight and output_norm.weight as the only non-block tensors and no separate output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 34 blocks, that yields five full-context global layers." }, { "kind": "sliding_window", "layers": 29, "kv_heads": 4, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 1024 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head, head dimension, and window metadata. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly selects a quantized KV cache." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The package also contains lower-bit GGUF siblings that require separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, image prefill, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches gemma-3-4b-it.fp16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant FP16 tensor format." }, "evidence": [ { "label": "MaziyarPanahi Gemma 3 4B IT GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-4b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 55ebd63de390fd82075d7ec5d1cf20486292d150, the live API records a public non-gated GGUF text-generation repo with base_model google/gemma-3-4b-it, base_model:quantized metadata, region:us, 142375 downloads, GGUF architecture gemma3, context_length 131072, gguf.total 3880099328, and gguf.totalFileSize 7767474368. The API totalFileSize matches gemma-3-4b-it.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Gemma 3 4B IT GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/raw/55ebd63de390fd82075d7ec5d1cf20486292d150/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model google/gemma-3-4b-it, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. It advertises multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a smaller default." }, { "label": "MaziyarPanahi Gemma 3 4B IT GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/raw/55ebd63de390fd82075d7ec5d1cf20486292d150/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral, which conflicts with the live API GGUF architecture and the selected GGUF header. This profile therefore does not use that config as architecture evidence." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 34 blocks recorded by the selected GGUF header." }, { "label": "Google Gemma 3 4B IT gated base config access check", "url": "https://huggingface.co/google/gemma-3-4b-it/raw/093f9f388b31de276ce2de164bdc2081324b9767/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The raw base config returned HTTP 401 with an access-restricted response in this audit environment. The existing google/gemma-3-4b-it profile is therefore still an unsupported gated-base stub, and this GGUF profile does not claim a direct base-config comparison." }, { "label": "MaziyarPanahi Gemma 3 4B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/tree/55ebd63de390fd82075d7ec5d1cf20486292d150", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 7.767474368 GB, Q8_0 4.130226368 GB, Q4_K_M 2.489757888 GB, and Q2_K 1.729028288 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Gemma 3 4B IT FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/resolve/55ebd63de390fd82075d7ec5d1cf20486292d150/gemma-3-4b-it.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 40 metadata entries and 444 tensors. The linked file is 7.767474368 GB. Tensor spans sum to 7.760934912 GB: token_embd.weight 1.342177280 GB, blk.* tensors 6.418747392 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006539456 GB. Stored tensor spans split into F16 7.759462400 GB and F32 0.001472512 GB. The header records general.architecture gemma3, gemma3.block_count 34, context_length 131072, embedding_length 2560, feed_forward_length 10240, attention.head_count 8, attention.head_count_kv 4, attention.key_length 256, attention.value_length 256, attention.sliding_window 1024, rope scaling factor 8, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card metadata, package config conflict check, gated-base access check, linked-object HEAD checks, Google Gemma 3 local/global attention documentation, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Gemma 3 4B IT FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--intellect-2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/INTELLECT-2-GGUF", "title": "MaziyarPanahi INTELLECT-2 GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of INTELLECT-2.", "model_family": "qwen2-32b-intellect-2-gguf", "base_model_proof": { "base_model": "PrimeIntellect/INTELLECT-2", "relation": "quantized", "source": "Hugging Face model card/API metadata, pinned base config/API metadata, package config absence check, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of PrimeIntellect/INTELLECT-2. The package does not publish a config.json, so the pinned base config and selected GGUF header are used for architecture evidence. Both record Qwen2 text geometry with 64 layers, hidden size 5120, 40 attention heads, 8 KV heads, and 40960 context." }, "architecture": { "canonical_architecture_id": "qwen2-32b-intellect-2", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 12.313098944, "swept_weight_gb": 12.051652608, "auxiliary_resident_weight_gb": 0.261446336, "resident_parameter_scope": "selected INTELLECT-2.Q2_K.gguf linked file size including GGUF metadata, tokenizer, header, and tensor payloads", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output.weight, and output_norm.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and alignment overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 12.313098944 GB. GGUF tensor payloads total 12.307120128 GB, while GGUF metadata, tokenizer, header, and alignment overhead account for 0.005978816 GB. The main GGUF stores separate token_embd.weight and output.weight tensors, so token_embd.weight is resident-only for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The pinned base config records use_sliding_window false, so all 64 Qwen2 layers use full-context attention despite the stale sliding_window field. Qwen2 attention caches separate key_states and value_states through DynamicCache with 8 KV heads and 128-dimensional heads." }, "notes": "This profile models ordinary text decode for the API-selected INTELLECT-2.Q2_K.gguf artifact. The larger Q3/Q4/Q5/Q6 siblings require separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3758132527333987, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor payloads for ordinary swept weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, activation traffic, and compute throughput are outside Bounds Engine v1.", "notes": "The selected artifact is INTELLECT-2.Q2_K.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "INTELLECT-2-GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/INTELLECT-2-GGUF", "source_type": "model_card", "supports": [ "repo", "pipeline", "revision", "downloads", "base_model_proof", "selected_artifact", "total_params_b" ], "notes": "At commit 276790213c9cb201745d2e85061e6e374b2ee9f7, the API records a public non-gated GGUF text-generation repo with base_model PrimeIntellect/INTELLECT-2, base_model:quantized metadata, region:us, conversational metadata, 110740 downloads, GGUF architecture qwen2, context_length 40960, gguf.total 32763876352, and gguf.totalFileSize 12313098944. The API totalFileSize matches INTELLECT-2.Q2_K.gguf, so this profile targets that artifact." }, { "label": "INTELLECT-2-GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF/raw/276790213c9cb201745d2e85061e6e374b2ee9f7/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "pipeline" ], "notes": "The package card records base_model PrimeIntellect/INTELLECT-2, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. It advertises multiple 2-bit through 8-bit quantizations and does not override the API-selected Q2_K artifact with a larger default." }, { "label": "PrimeIntellect INTELLECT-2 API metadata", "url": "https://huggingface.co/api/models/PrimeIntellect/INTELLECT-2", "source_type": "model_card", "supports": [ "license", "base_model_proof", "total_params_b", "serving" ], "notes": "At commit 93e0fc8da282b7273d8d432878b626843d97d667, the base API records a public non-gated Apache-2.0 safetensors Qwen2 repo with arXiv and region:us metadata. The safetensors block records F32 32763876352 and total 32763876352 parameters, matching the selected GGUF logical total." }, { "label": "PrimeIntellect INTELLECT-2 config", "url": "https://huggingface.co/PrimeIntellect/INTELLECT-2/raw/93e0fc8da282b7273d8d432878b626843d97d667/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The base config records Qwen2ForCausalLM, model_type qwen2, torch_dtype float32, hidden_size 5120, intermediate_size 27648, 64 hidden layers, 40 attention heads, 8 KV heads, 40960 max position embeddings, rope_theta 1000000, tie_word_embeddings false, use_cache true, use_sliding_window false, sliding_window 32768, and vocab_size 152064." }, { "label": "INTELLECT-2-GGUF package config absence check", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF/raw/276790213c9cb201745d2e85061e6e374b2ee9f7/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package siblings list does not include config.json, and a raw config request returned HTTP 404. This profile therefore uses the selected GGUF header and pinned base config as architecture evidence." }, { "label": "Transformers Qwen2 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/08a7ef05bcf9723cb2e58855afb8dc2c799323ff/src/transformers/models/qwen2/modeling_qwen2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "ordinary_decode_scope" ], "notes": "Manual review found Qwen2Attention builds q_proj, k_proj, v_proj, and o_proj, caches key_states and value_states with DynamicCache, and only applies sliding_window when config.layer_types marks a layer as sliding_attention. Qwen2Config sets sliding_window to null when use_sliding_window is false, making this INTELLECT-2 config full-context across all layers." }, { "label": "INTELLECT-2-GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF/tree/276790213c9cb201745d2e85061e6e374b2ee9f7", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Exact HEAD checks found Q2_K 12.313098944 GB, Q3_K_M 15.935048384 GB, Q3_K_L 17.247079104 GB, Q4_K_M 19.851336384 GB, Q5_K_M 23.262157504 GB, and Q6_K 26.886154944 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "INTELLECT-2 Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF/resolve/276790213c9cb201745d2e85061e6e374b2ee9f7/INTELLECT-2.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 31 metadata entries and 771 tensors. The linked file is 12.313098944 GB. Tensor payloads sum to 12.307120128 GB across 32.763876352B logical elements: output.weight 0.638668800 GB, token_embd.weight 0.255467520 GB, blk.* tensors 11.412963328 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.005978816 GB. Stored tensor payloads split into Q2_K 6.861496320 GB, Q3_K 4.613734400 GB, Q6_K 0.638668800 GB, Q4_K 0.188743680 GB, and F32 0.004476928 GB. The header records general.architecture qwen2, Apache-2.0 license, qwen2.block_count 64, context_length 40960, embedding_length 5120, feed_forward_length 27648, attention.head_count 40, attention.head_count_kv 8, rope.freq_base 1000000, vocab_size 152064, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned package and base model cards, pinned base config/API metadata, package config absence check, Transformers Qwen2 implementation review, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected Q2_K artifact." }, "notes": "Use this profile for the API-selected MaziyarPanahi INTELLECT-2 Q2_K GGUF artifact. Do not infer the larger sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--llama-3-2-1b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF", "title": "MaziyarPanahi Llama 3.2 1B Instruct GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Llama 3.2 1B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-1B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, and GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.2-1B-Instruct as the quantized base. The repo-local config.json is only a 31-byte model_type stub that incorrectly says mistral, and the base raw config is gated in this audit environment, so this profile uses the selected public GGUF header as the direct architecture source." }, "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.235814432, "swept_params_b": 1.235814432, "resident_weight_gb": 2.479595392, "swept_weight_gb": 2.471764096, "auxiliary_resident_weight_gb": 0.007831296, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.2-1B-Instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected F16 linked file is 2.479595392 GB. Header tensor spans total 2.471764096 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007831296 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 16-layer decoder with 8 KV heads and 64-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. It uses the public GGUF metadata for architecture because the Meta base config is gated and the repo-local config stub is not architecture evidence." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0064463788362685, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains many GGUF quantizations. This profile intentionally targets Llama-3.2-1B-Instruct.fp16.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "MaziyarPanahi Llama 3.2 1B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit b64ae94264258a3d7516a34a8c54928d32b19869, the API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.2-1B-Instruct, region:us, 119953 downloads, GGUF architecture llama, 131072 context length, gguf.total 1235814432, and gguf.totalFileSize 2479595392." }, { "label": "MaziyarPanahi Llama 3.2 1B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "quantization_recipe" ], "notes": "The card records this as a GGUF repo by MaziyarPanahi for meta-llama/Llama-3.2-1B-Instruct and lists a multi-quant GGUF package." }, { "label": "MaziyarPanahi Llama 3.2 1B Instruct GGUF config stub", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF/raw/b64ae94264258a3d7516a34a8c54928d32b19869/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json is only 31 bytes and contains {\"model_type\":\"mistral\"}. The selected GGUF header and API gguf metadata identify the actual architecture as llama, so this stub is treated as stale package metadata and is not used for bounds geometry." }, { "label": "Llama 3.2 1B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 9213176726f574b556790deb65791e0c5aa438b6, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 1235814400 parameters." }, { "label": "Llama 3.2 1B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A direct raw config request returned 401 restricted access. The profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "MaziyarPanahi Llama 3.2 1B Instruct GGUF linked-file size metadata", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF/tree/b64ae94264258a3d7516a34a8c54928d32b19869", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata found GGUF siblings with sizes: IQ1_M 413605984, IQ1_S 393551968, IQ2_XS 475865184, IQ3_XS 621113440, IQ4_XS 743141472, Q2_K 580874336, Q3_K_L 732524640, Q3_K_M 690843744, Q3_K_S 641691744, Q4_K_M 807694432, Q4_K_S 775647328, Q5_K_M 911503456, Q5_K_S 892563552, Q6_K 1021800544, Q8_0 1321082976, and fp16 2479595392 bytes. The selected fp16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Llama 3.2 1B Instruct F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF/resolve/b64ae94264258a3d7516a34a8c54928d32b19869/Llama-3.2-1B-Instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 31 metadata entries and 147 tensors. The linked file is 2.479595392 GB. Tensor spans sum to 2.471764096 GB: token_embd.weight 0.525336576 GB, blk.* tensors 1.946419200 GB, output_norm.weight 0.000008192 GB, and rope_freqs.weight 0.000000128 GB. Metadata/tokenizer/header/file overhead accounts for 0.007831296 GB. Stored tensor bytes split into F16 2.471493632 GB and F32 0.000270464 GB. The header records llama.block_count 16, context_length 131072, embedding_length 2048, feed_forward_length 8192, attention.head_count 32, attention.head_count_kv 8, attention key/value length 64, rope.freq_base 500000, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, repo config-stub check, base-model API metadata, gated base-config access check, expanded linked-file size metadata, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Llama 3.2 1B Instruct F16 GGUF artifact. Do not use the repo-local config stub for architecture; the architecture evidence is the selected GGUF header metadata." }, { "id": "maziyarpanahi--llama-3-2-3b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF", "title": "MaziyarPanahi Llama 3.2 3B Instruct GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the row-selected Q2_K GGUF artifact of Llama 3.2 3B Instruct.", "model_family": "llama-3.2-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-3B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, base-model API metadata, gated base-config access check, stale repo-config check, and selected Q2_K GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of meta-llama/Llama-3.2-3B-Instruct. The base raw config is gated in this audit environment, and the repo-local config.json incorrectly says model_type mistral, so this profile uses the selected public Q2_K GGUF header as the direct architecture source." }, "architecture": { "canonical_architecture_id": "llama-3-2-3b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.212749888, "swept_params_b": 3.212749888, "auxiliary_resident_params_b": 0, "resident_weight_gb": 1.363935776, "swept_weight_gb": 1.356097792, "auxiliary_resident_weight_gb": 0.007837984, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.2-3B-Instruct.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "notes": "The selected Q2_K linked file is 1.363935776 GB. Header tensor spans total 1.356097792 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007837984 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and rope_freqs.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic. The selected artifact mixes Q2_K, Q3_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records a Llama-style 28-layer decoder with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the row-selected Q2_K GGUF artifact. It uses the public selected GGUF metadata for architecture because the Meta base config is gated and the repo-local config is stale." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.4245384245734351, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, tokenizer processing, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets Llama-3.2-3B-Instruct.Q2_K.gguf because the catalog row is the 2-bit GGUF entry. The live HF API gguf.totalFileSize currently matches the FP16 sibling instead, so exact selected-file header bytes are authoritative. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Llama 3.2 3B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit e56a0ae870579697698c3ded68df97747125d554, the API records a public non-gated GGUF text-generation repo with base_model meta-llama/Llama-3.2-3B-Instruct, base_model:quantized metadata, quantized tags, region:us, 119553 current downloads, GGUF architecture llama, context_length 131072, gguf.total 3212749888, and gguf.totalFileSize 6433687872. The API totalFileSize matches the FP16 sibling, while this profile targets the Q2_K file for the catalog's 2-bit row." }, { "label": "MaziyarPanahi Llama 3.2 3B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/raw/e56a0ae870579697698c3ded68df97747125d554/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The pinned card records original model meta-llama/Llama-3.2-3B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base-model API record Llama 3.2 licensing." }, { "label": "Llama 3.2 3B Instruct base-model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "At commit 0cb88a4f764b7a12671c53f0838cd831a0843b95, the base-model API records a gated-manual Transformers Llama text-generation repo with Llama 3.2 license, region:us tag, and BF16 safetensors total 3212749824 parameters." }, { "label": "Llama 3.2 3B Instruct gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/raw/0cb88a4f764b7a12671c53f0838cd831a0843b95/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A direct raw config request returned HTTP 401. This GGUF profile therefore does not infer layer count, KV heads, context length, or tied embedding layout from the gated base config." }, { "label": "MaziyarPanahi Llama 3.2 3B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/tree/e56a0ae870579697698c3ded68df97747125d554", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 1.363935776 GB, Q3_K_S 1.542849056 GB, Q3_K_M 1.687159328 GB, Q3_K_L 1.815347744 GB, IQ4_XS 1.829110304 GB, Q4_K_S 1.928200736 GB, Q4_K_M 2.019377696 GB, Q5_K_S 2.269512224 GB, Q5_K_M 2.322154016 GB, Q6_K 2.643853856 GB, Q8_0 3.421899296 GB, and fp16 6.433687872 GB. The selected Q2_K artifact is the standard 2-bit GGUF file for this row; the API gguf.totalFileSize points to the fp16 sibling instead." }, { "label": "MaziyarPanahi Llama 3.2 3B Instruct Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/resolve/e56a0ae870579697698c3ded68df97747125d554/Llama-3.2-3B-Instruct.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 255 tensors. The linked file is 1.363935776 GB. Tensor spans sum to 1.356097792 GB: token_embd.weight 0.323205120 GB, blk.* tensors 1.032880128 GB, output_norm.weight 0.000012288 GB, and rope_freqs.weight 0.000000256 GB. Metadata/tokenizer/header/file overhead accounts for 0.007837984 GB. Stored tensor spans split into Q2_K 0.578027520 GB, Q3_K 0.454164480 GB, Q6_K 0.323205120 GB, and F32 0.000700672 GB. The header records general.architecture llama, Llama 3.2 license, llama.block_count 28, context_length 131072, embedding_length 3072, feed_forward_length 8192, attention.head_count 24, attention.head_count_kv 8, key/value length 128, rope.freq_base 500000, and no separate output.weight tensor." }, { "label": "MaziyarPanahi Llama 3.2 3B Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/raw/e56a0ae870579697698c3ded68df97747125d554/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base-model metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, base-model API metadata, gated base-config access check, stale repo-config check, linked GGUF file sizes, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the row-selected Llama 3.2 3B Instruct Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--llama-3-3-70b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF", "title": "MaziyarPanahi Llama 3.3 70B Instruct GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Llama 3.3 70B Instruct.", "model_family": "llama3.3-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.3-70B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, gated base-config access check, and audited Casper Hansen Llama 3.3 70B AWQ config comparison", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of meta-llama/Llama-3.3-70B-Instruct. The base repo remains gated in this audit environment and this GGUF repo has no config.json, so direct base-config compatibility cannot be independently verified. The selected Q2_K GGUF header records the same served Llama 3.3 geometry as the audited Casper Hansen AWQ profile: 80 layers, hidden size 8192, intermediate size 28672, 64 attention heads, 8 KV heads, 128 key/value head dimension, 131072 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "llama-3-3-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.55370656, "swept_params_b": 69.503033408, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 26.37511344, "swept_weight_gb": 26.022494464, "auxiliary_resident_weight_gb": 0.352618976, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3.3-70B-Instruct.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, rope_freqs.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 26.375113440 GB. GGUF tensor spans total 26.367246592 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007866848 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q5_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 80 Llama decoder layers, 8 KV heads, 128-dimensional key/value heads, and 131072 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. Split Q6_K, Q8_0, FP16, and larger single-file quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.373830302134023, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the Q2_K GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Llama 3.3 70B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit b09bf396ffd1df6c4c46d66c932f2c3cc2b7b21c, the API records a public non-gated GGUF text-generation repo with base_model meta-llama/Llama-3.3-70B-Instruct, base_model:quantized metadata, quantized tags, region:us, 125907 downloads, GGUF architecture llama, context_length 131072, gguf.total 70553706560, and gguf.totalFileSize 26375113440." }, { "label": "MaziyarPanahi Llama 3.3 70B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model meta-llama/Llama-3.3-70B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base-model API record Llama 3.3 licensing." }, { "label": "Meta Llama 3.3 70B Instruct gated base profile", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/raw/6f6073b423013f6a7d4d9f39144961bfbfbc386b/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The existing base profile records that raw config, README, model.safetensors.index.json, and hf download config.json requests require manual approval in this audit environment. This GGUF profile therefore does not infer architecture from the gated base config." }, { "label": "Casper Hansen Llama 3.3 70B AWQ audited profile", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "The audited AWQ profile records a public served config for a Llama 3.3 70B Instruct derivative with LlamaForCausalLM, hidden size 8192, intermediate size 28672, 80 layers, 64 attention heads, 8 KV heads, head_dim 128, 131072 context, tie_word_embeddings false, and separate input/output embeddings. The selected GGUF header matches that geometry." }, { "label": "MaziyarPanahi Llama 3.3 70B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF/tree/b09bf396ffd1df6c4c46d66c932f2c3cc2b7b21c", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found Q2_K 26.375113440 GB, Q3_K_L 37.140597472 GB, Q3_K_M 34.267499232 GB, Q3_K_S 30.912056032 GB, Q4_K_M 42.520398560 GB, Q4_K_S 40.347224800 GB, Q5_K_M 49.949821664 GB, Q5_K_S 48.657451744 GB, split Q6_K total 57.888148896 GB, split Q8_0 total 74.975055264 GB, and split FP16 total 141.117918624 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Llama 3.3 70B Instruct Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF/resolve/b09bf396ffd1df6c4c46d66c932f2c3cc2b7b21c/Llama-3.3-70B-Instruct.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 36 metadata entries and 724 tensors. The linked file is 26.375113440 GB. Tensor spans sum to 26.367246592 GB: output.weight 0.861880320 GB, token_embd.weight 0.344752128 GB, blk.* tensors 25.160581120 GB, output_norm.weight 0.000032768 GB, and rope_freqs.weight 0.000000256 GB. Metadata/tokenizer/header/file overhead accounts for 0.007866848 GB. Stored tensor bytes split into Q2_K 14.657814528 GB, Q3_K 10.380902400 GB, Q5_K 0.461373440 GB, Q6_K 0.861880320 GB, and F32 0.005275904 GB. The header records general.architecture llama, Llama 3.3 license, llama.block_count 80, context_length 131072, embedding_length 8192, feed_forward_length 28672, attention.head_count 64, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 500000, key/value length 128, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Llama 3.3 70B Instruct GGUF missing config check", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF/raw/b09bf396ffd1df6c4c46d66c932f2c3cc2b7b21c/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json returns 404, and the API config object is empty. This profile intentionally uses the selected GGUF header as the architecture source of truth." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, gated base profile, audited Casper Hansen AWQ config comparison, expanded linked-file size metadata for all GGUF siblings, missing repo-config check, and a direct GGUF header/tensor-index range read of the selected Q2_K artifact." }, "notes": "Use this profile for the API-selected Llama 3.3 70B Instruct Q2_K GGUF artifact. Do not infer the lower-bit or split-file sibling footprints from this profile." }, { "id": "maziyarpanahi--llama-3-8b-instruct-32k-v0-1-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF", "title": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Llama 3 8B Instruct 32k v0.1.", "model_family": "llama3-dense-gguf", "base_model_proof": { "base_model": "MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1", "relation": "quantized", "source": "Hugging Face model card/API metadata, inaccessible base API check, package config conflict check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1. The base repo API returned 401 in this audit environment, and the package config.json incorrectly records only model_type mistral, so this profile uses the selected public GGUF header as the architecture source instead of copying or inferring from the inaccessible base config." }, "architecture": { "canonical_architecture_id": "llama-3-8b-32k-v0-1", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 16.068890944, "swept_weight_gb": 15.010381824, "auxiliary_resident_weight_gb": 1.05850912, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3-8B-Instruct-32k-v0.1.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices for each generated token", "notes": "The selected FP16 linked file is 16.068890944 GB. GGUF tensor spans total 16.061054976 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.007835968 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 decoder layers, 8 KV heads, 128-dimensional key/value heads, and 8192 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving. The repo name says 32k, but both the live API GGUF metadata and selected artifact header record 8192 context tokens." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The repo contains many lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches Llama-3-8B-Instruct-32k-v0.1.fp16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant FP16 tensor format." }, "evidence": [ { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 4251e4b7ec372ba0de6046137272a709e349f442, the live API records a public non-gated GGUF text-generation repo with base_model MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1, region:us, 122730 downloads, GGUF architecture llama, context_length 8192, gguf.total 8030261248, and gguf.totalFileSize 16068890944. The API totalFileSize matches Llama-3-8B-Instruct-32k-v0.1.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/raw/4251e4b7ec372ba0de6046137272a709e349f442/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. The card advertises multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a smaller default." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/raw/4251e4b7ec372ba0de6046137272a709e349f442/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral, which conflicts with the live API GGUF architecture and the selected GGUF header. This profile therefore does not use that config as architecture evidence." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 base API access check", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A live base API request returned 401 with Invalid username or password in this audit environment. This GGUF profile therefore does not claim a direct base-config comparison, but it can use the public selected GGUF header for the served GGUF artifact's exact geometry." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/tree/4251e4b7ec372ba0de6046137272a709e349f442", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 16.068890944 GB, Q8_0 8.540770816 GB, Q6_K 6.596006400 GB, Q5_K_M 5.732987392 GB, Q4_K_M 4.920734208 GB, Q3_K_M 4.018917888 GB, Q2_K 3.179131392 GB, IQ4_XS 4.447662592 GB, IQ3_XS 3.518747136 GB, IQ2_XS 2.605781504 GB, IQ1_M 2.161971712 GB, and IQ1_S 2.019627520 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 32k v0.1 FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/resolve/4251e4b7ec372ba0de6046137272a709e349f442/Llama-3-8B-Instruct-32k-v0.1.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 22 metadata entries and 291 tensors. The linked file is 16.068890944 GB. Tensor spans sum to 16.061054976 GB: token_embd.weight 1.050673152 GB, output.weight 1.050673152 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.007835968 GB. Stored tensor spans split into F16 16.059990016 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 8192, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 8000000, vocab_size 128256, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, package config conflict check, inaccessible base API check, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Llama 3 8B Instruct 32k v0.1 FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--llama-3-8b-instruct-64k-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF", "title": "MaziyarPanahi Llama 3 8B Instruct 64k GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Llama 3 8B Instruct 64k.", "model_family": "llama3-8b-64k-dense-gguf", "base_model_proof": { "base_model": "MaziyarPanahi/Llama-3-8B-Instruct-64k", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object size checks, base model card/config, and direct selected FP16 GGUF tensor-index range read", "config_compatible": false, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of MaziyarPanahi/Llama-3-8B-Instruct-64k. The selected GGUF header matches the base Llama 3 layer, hidden, feed-forward, head, KV-head, vocabulary, RoPE theta, and untied embedding geometry, but the base Transformers config still records max_position_embeddings 8192 while the base card and selected GGUF header record 64000-token GGUF context. This profile therefore uses the selected GGUF header as serving truth and records the base-config context mismatch explicitly." }, "architecture": { "canonical_architecture_id": "llama-3-8b-instruct-64k-gguf-text", "max_context_tokens": 64000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 16.068890848, "swept_weight_gb": 15.010381824, "auxiliary_resident_weight_gb": 1.058509024, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Llama-3-8B-Instruct-64k.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 16.068890848 GB. GGUF tensor payloads total 16.061054976 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007835872 GB. Because output.weight is stored separately and the base config records tie_word_embeddings false, token_embd.weight is resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records llama.context_length 64000, 32 layers, 8 KV heads, 32 attention heads, 4096 hidden size, and 128-dimensional RoPE/head geometry. It contains no sliding-window metadata, so this selected-artifact profile charges full-context FP16 K/V for llama.cpp-style serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. It does not substitute lower-bit GGUF siblings." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Llama-3-8B-Instruct-64k.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Llama 3 8B Instruct 64k GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 3785b79d1c23332e60859b853ed9807f94844353, the API records a public non-gated GGUF repo with base_model MaziyarPanahi/Llama-3-8B-Instruct-64k, quantized tags, llama-3 tags, region:us, text-generation pipeline, 113638 downloads, GGUF architecture llama, context length 64000, gguf.total 8030261248, and gguf.totalFileSize 16068890848. The API totalFileSize matches Llama-3-8B-Instruct-64k.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF/raw/3785b79d1c23332e60859b853ed9807f94844353/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The GGUF card records base_model MaziyarPanahi/Llama-3-8B-Instruct-64k, model_creator MaziyarPanahi, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists multiple lower-bit siblings, but the API-selected totalFileSize points at the FP16 sibling." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k base API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-64k", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit 7423c85617da4b946d251274c6958d9533fbbc5c, the base repo is public, Llama 3 licensed, references winglian/Llama-3-8b-64k-PoSE as base_model, and records F16 safetensors total 8030261248 parameters." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k base model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k/raw/7423c85617da4b946d251274c6958d9533fbbc5c/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "max_context_tokens", "context_mismatch" ], "notes": "The base card says the model is based on winglian/Llama-3-8b-64k-PoSE, uses PoSE to extend Llama context from 8k to 64k, and explicitly states that all GGUF models come with context length 64000." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k base config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k/raw/7423c85617da4b946d251274c6958d9533fbbc5c/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "context_mismatch" ], "notes": "The base config records LlamaForCausalLM, model_type llama, FP16 dtype, 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, rope_theta 500000, tie_word_embeddings false, and vocab_size 128256. It still records max_position_embeddings 8192, which conflicts with the base card's 64k GGUF statement and the selected GGUF header's context_length 64000." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF/tree/3785b79d1c23332e60859b853ed9807f94844353", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata found GGUF siblings Q2_K 3.179131136 GB, Q3_K_L 4.321956096 GB, Q3_K_M 4.018917632 GB, Q3_K_S 3.664498944 GB, Q4_K_M 4.920733952 GB, Q4_K_S 4.692668672 GB, Q5_K_M 5.732987136 GB, Q5_K_S 5.599293696 GB, Q6_K 6.596006144 GB, Q8_0 8.540770560 GB, and FP16 16.068890848 GB. The FP16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Llama 3 8B Instruct 64k FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF/resolve/3785b79d1c23332e60859b853ed9807f94844353/Llama-3-8B-Instruct-64k.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 20 metadata entries and 291 tensors. The selected file is 16.068890848 GB, with tensor payloads starting at byte 7835872. Tensor payloads sum to 16.061054976 GB across 8.030261248B logical elements: token_embd.weight 1.050673152 GB, output.weight 1.050673152 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.007835872 GB. Stored tensor bytes split into F16 16.059990016 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 64000, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 500000, vocab_size 128256, and no llama.attention.sliding_window field." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model cards, package config-stub check, base API/config metadata, selected linked-object size checks, and direct selected FP16 GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected FP16 GGUF artifact. Do not silently substitute lower-bit siblings." }, { "id": "maziyarpanahi--mathstral-7b-v0-1-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/mathstral-7B-v0.1-GGUF", "title": "MaziyarPanahi Mathstral 7B v0.1 GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Mathstral 7B v0.1.", "model_family": "mathstral-mistral-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mathstral-7B-v0.1", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, selected linked-object size checks, and Mathstral base config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mistralai/Mathstral-7B-v0.1. The base config records MistralForCausalLM geometry with 32 layers, 32 attention heads, 8 KV heads, 128 head dimension, 4096 hidden size, 14336 intermediate size, untied embeddings, no sliding window, and 32768 max positions. The selected GGUF header records llama.cpp architecture llama with matching Mistral-compatible layer, attention, KV, head-dimension, RoPE, context, vocabulary, and untied embedding geometry. The repo-local package config is only a stale model_type mistral stub and is not used as the architecture source." }, "architecture": { "canonical_architecture_id": "mathstral-7b-v0-1", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.248023552, "swept_params_b": 7.113805824, "auxiliary_resident_params_b": 0.134217728, "resident_weight_gb": 14.497336928, "swept_weight_gb": 14.228144128, "auxiliary_resident_weight_gb": 0.2691928, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for mathstral-7B-v0.1.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 14.497336928 GB. GGUF tensor payloads total 14.496579584 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000757344 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 32768 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is mathstral-7B-v0.1.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mathstral 7B v0.1 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/mathstral-7B-v0.1-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit d3c8fafa377cf389d822360b1a289226ba904d7d, the API records a public non-gated GGUF repo with base_model mistralai/Mathstral-7B-v0.1, quantized tags, region:us, text-generation pipeline, 113042 downloads, GGUF architecture llama, context length 32768, gguf.total 7248023552, and gguf.totalFileSize 14497336928. The API totalFileSize matches mathstral-7B-v0.1.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mathstral 7B v0.1 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/raw/d3c8fafa377cf389d822360b1a289226ba904d7d/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The model card records base_model mistralai/mathstral-7B-v0.1, model_creator mistralai, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists many quantized siblings, but the API-selected totalFileSize points at the F16 sibling." }, { "label": "Mathstral 7B v0.1 base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mathstral-7B-v0.1", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit ec3a48484ef241dfe03282edcb0f25e564923823, the base repo is public, Apache-2.0 licensed, and records F32 safetensors parameters 7248023552." }, { "label": "Mathstral 7B v0.1 base config", "url": "https://huggingface.co/mistralai/Mathstral-7B-v0.1/raw/ec3a48484ef241dfe03282edcb0f25e564923823/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 32768, sliding_window null, rope_theta 1000000, vocab size 32768, and tie_word_embeddings false." }, { "label": "Mathstral 7B v0.1 params.json", "url": "https://huggingface.co/mistralai/Mathstral-7B-v0.1/raw/ec3a48484ef241dfe03282edcb0f25e564923823/params.json", "source_type": "config", "supports": [ "architecture", "kv_adapter" ], "notes": "The Mistral params.json independently records dim 4096, 32 layers, 32 heads, 8 KV heads, head_dim 128, hidden_dim 14336, vocab_size 32768, and rope_theta 1000000." }, { "label": "MaziyarPanahi Mathstral 7B v0.1 GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/tree/d3c8fafa377cf389d822360b1a289226ba904d7d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ1_S 1.615318848 GB, IQ1_M 1.757663040 GB, IQ2_XS 2.201472832 GB, IQ3_XS 3.022769984 GB, IQ4_XS 3.911962432 GB, Q2_K 2.722877248 GB, Q3_K_S 3.168522048 GB, Q3_K_M 3.522940736 GB, Q3_K_L 3.825979200 GB, Q4_K_S 4.144746304 GB, Q4_K_M 4.372811584 GB, Q5_K_S 5.002481472 GB, Q5_K_M 5.136174912 GB, Q6_K 5.947248448 GB, Q8_0 7.702564672 GB, and fp16 14.497336928 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mathstral 7B v0.1 F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/resolve/d3c8fafa377cf389d822360b1a289226ba904d7d/mathstral-7B-v0.1.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 26 metadata entries and 291 tensors. The selected file is 14.497336928 GB, with tensor payloads starting at byte 757344. Tensor payloads sum to 14.496579584 GB across 7.248023552B logical elements: token_embd.weight 0.268435456 GB, output.weight 0.268435456 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.000757344 GB. Stored tensor bytes split into F16 14.495514624 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 32768, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000, and vocab_size 32768." }, { "label": "MaziyarPanahi Mathstral 7B v0.1 GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/raw/d3c8fafa377cf389d822360b1a289226ba904d7d/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral. It is recorded as a stale package stub and is not used as the main architecture evidence because the selected GGUF header and pinned base config provide the full served geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, base Mathstral API/config metadata, params.json, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--meta-llama-3-1-405b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF", "title": "MaziyarPanahi Meta Llama 3.1 405B Instruct GGUF Q2_K Split", "summary": "Audited memory-side text-decode bounds profile for the API-selected 9-part Q2_K split GGUF artifact of Meta Llama 3.1 405B Instruct.", "model_family": "llama31-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-405B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected split-GGUF header metadata, and selected linked-object size checks", "config_compatible": false, "notes": "The GGUF repo card and API metadata identify this package as a quantized derivative of meta-llama/Llama-3.1-405B-Instruct. The Meta base repo is gated in this audit environment, so this profile uses the public selected split-GGUF headers as the architecture source instead of claiming a direct base-config comparison." }, "architecture": { "canonical_architecture_id": "llama-3-1-405b-instruct", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 410.081247232, "swept_params_b": 407.979900928, "auxiliary_resident_params_b": 2.101346304, "resident_weight_gb": 151.205502336, "swept_weight_gb": 150.508109824, "auxiliary_resident_weight_gb": 0.697392512, "resident_parameter_scope": "selected 9-part Q2_K split GGUF linked-file size and API GGUF logical tensor parameters", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K split GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, headers, and file overhead are resident in the selected split artifact but not swept for ordinary text decode", "notes": "HF API gguf.total is 410.081247232B parameters and gguf.totalFileSize selects the 9-part Meta-Llama-3.1-405B-Instruct.Q2_K split. Range-reads of all nine GGUF v3 shard headers found 1137 tensors. Linked split files total 151.205502336 GB. Tensor spans total 151.197614080 GB, while metadata/header/tokenizer/alignment overhead accounts for 0.007888256 GB. token_embd.weight is 0.689504256 GB and is resident-only because output.weight is stored separately." }, "kv_adapter": { "kind": "full_context", "layers": 126, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected split-GGUF header records llama.block_count 126, llama.attention.head_count 128, llama.attention.head_count_kv 16, llama.rope.dimension_count 128, and no sliding-window metadata, so this selected-artifact profile charges full-context FP16 K/V for llama.cpp-style serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K split GGUF artifact. It does not substitute the larger Q3_K_S split." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3687208409470545, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-split-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected split-GGUF tensor spans for weight traffic and selected linked-file bytes for residency. GGUF split loading overhead, kernels, dequantization, scheduler behavior, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The API-selected artifact is the 9-part Q2_K split because the split linked-file sum exactly matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Meta Llama 3.1 405B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 85b9bd67025a4337e9694ec0edaf46437fe6283b, the live API records a public non-gated GGUF repo with base_model meta-llama/Llama-3.1-405B-Instruct, base_model_relation quantized, Llama 3.1 license metadata, region:us, 108344 downloads, GGUF architecture llama, context length 131072, gguf.total 410081247232, and gguf.totalFileSize 151205502336." }, { "label": "MaziyarPanahi Meta Llama 3.1 405B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF/raw/85b9bd67025a4337e9694ec0edaf46437fe6283b/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The card metadata records base_model meta-llama/Llama-3.1-405B-Instruct, quantized_by MaziyarPanahi, GGUF/llama.cpp packaging, Llama 3.1 licensing, supported languages, and available Q2_K and Q3_K_S split artifacts." }, { "label": "Meta Llama 3.1 405B Instruct gated base access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct/raw/be673f326cab4cd22ccfef76109faf68e41aa5f1/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "config_access" ], "notes": "The upstream Meta repo remains manually gated in this audit environment. Existing access checks for the base config and tensor index returned gated access, so this GGUF profile does not claim direct base-config compatibility." }, { "label": "MaziyarPanahi Meta Llama 3.1 405B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF/tree/85b9bd67025a4337e9694ec0edaf46437fe6283b", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Exact HEAD checks found the nine Q2_K split shards total 151.205502336 GB, exactly matching API gguf.totalFileSize. The alternate nine Q3_K_S split shards total 177.050312064 GB and are not selected by this profile." }, { "label": "MaziyarPanahi Meta Llama 3.1 405B Instruct Q2_K split GGUF header audit", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF/resolve/85b9bd67025a4337e9694ec0edaf46437fe6283b/Meta-Llama-3.1-405B-Instruct.Q2_K.gguf-00001-of-00009.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "Range-reads of the selected Q2_K split GGUF v3 headers found split.count 9 and split.tensors.count 1137. Header tensor spans across all nine shards sum to 151.197614080 GB: blk.* tensors 148.784283648 GB, output.weight 1.723760640 GB, token_embd.weight 0.689504256 GB, and output_norm.weight 0.000065536 GB. Metadata/header/tokenizer/alignment overhead across the split accounts for 0.007888256 GB. Stored tensor spans split into Q2_K 85.312733184 GB, Q3_K 61.766369280 GB, Q4_K 2.378170368 GB, Q6_K 1.723760640 GB, and F32 0.016580608 GB. The first shard header records llama.block_count 126, context_length 131072, embedding_length 16384, feed_forward_length 53248, attention.head_count 128, attention.head_count_kv 16, rope.dimension_count 128, rope.freq_base 500000, vocab_size 128256, and no sliding-window metadata." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, HF CLI model info, model card metadata, gated-base access context, linked-object HEAD checks, and direct GGUF header/tensor-index range reads of all nine selected Q2_K split shards." }, "notes": "Use this profile for the API-selected Q2_K split GGUF artifact. Do not infer the Q3_K_S split footprint unless the workload profile explicitly selects and audits that split." }, { "id": "maziyarpanahi--meta-llama-3-1-70b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF", "title": "MaziyarPanahi Meta Llama 3.1 70B Instruct GGUF IQ1_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected IQ1_M GGUF artifact of Meta Llama 3.1 70B Instruct.", "model_family": "llama3.1-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-70B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, gated base API/access checks, and audited Hugging Quants Llama 3.1 70B AWQ config comparison", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of the gated Meta Llama 3.1 70B Instruct model. The cardData/frontmatter use the historical Meta-Llama spelling, while the API resolves that alias to meta-llama/Llama-3.1-70B-Instruct. The gated base repo still blocks raw config and tensor-index access in this audit environment, and this GGUF repo has no config.json, so direct base-config compatibility cannot be independently verified. The selected IQ1_M GGUF header records the same served Llama 3.1 geometry as the audited Hugging Quants AWQ profile: 80 layers, hidden size 8192, intermediate size 28672, 64 attention heads, 8 KV heads, 128 RoPE dimension, 131072 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "llama-3-1-70b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 70.553706496, "swept_params_b": 69.503033344, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 16.751196288, "swept_weight_gb": 16.39858176, "auxiliary_resident_weight_gb": 0.352614528, "resident_parameter_scope": "selected GGUF linked file size and API/header logical tensor parameters for Meta-Llama-3.1-70B-Instruct.IQ1_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected IQ1_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected IQ1_M linked file is 16.751196288 GB. GGUF tensor spans total 16.743333888 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007862400 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes IQ1_M, IQ2_XXS, Q2_K, Q4_K, Q5_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 1-bit or 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 80 Llama decoder layers, 8 KV heads, 128-dimensional RoPE/KV heads, and 131072 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected IQ1_M GGUF artifact. Q2_K, Q3_K, Q4_K, Q5_K, and split Q6_K siblings should get separate profiles if explicitly selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.23742475229064963, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq1-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the IQ1_M GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Meta Llama 3.1 70B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit c1c0c3d88f2fca6f18de9063917f9bc49a33365a, the API records a public non-gated GGUF text-generation repo with base_model metadata for Meta Llama 3.1 70B Instruct, base_model:quantized metadata, quantized tags, region:us, 122003 downloads, GGUF architecture llama, context_length 131072, gguf.total 70553706496, and gguf.totalFileSize 16751196288. The totalFileSize matches Meta-Llama-3.1-70B-Instruct.IQ1_M.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Meta Llama 3.1 70B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF/raw/c1c0c3d88f2fca6f18de9063917f9bc49a33365a/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model meta-llama/Meta-Llama-3.1-70B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It advertises multiple lower-bit quantizations but does not override the API-selected IQ1_M artifact with a different default." }, { "label": "Meta Llama 3.1 70B Instruct gated base API and access checks", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-70B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "logical_parameter_count", "license" ], "notes": "The current base API records gated manual access at commit 1605565b47bb9346c5515c34102e054115b4f98b, text-generation pipeline, Llama 3.1 license, region:us, and BF16 safetensors total 70553706496 parameters. Raw config, README, and model.safetensors.index.json requests still return restricted-access 401 responses in this audit environment." }, { "label": "Hugging Quants Llama 3.1 70B AWQ audited profile", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "The audited AWQ profile records a public served config for a Llama 3.1 70B Instruct derivative with LlamaForCausalLM, hidden size 8192, intermediate size 28672, 80 layers, 64 attention heads, 8 KV heads, max_position_embeddings 131072, tie_word_embeddings false, and separate input/output embeddings. The selected GGUF header matches that geometry." }, { "label": "MaziyarPanahi Meta Llama 3.1 70B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF/tree/c1c0c3d88f2fca6f18de9063917f9bc49a33365a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ1_M 16.751196288 GB, IQ1_S 15.343483008 GB, IQ2_XS 21.142108288 GB, IQ3_XS 29.307730048 GB, IQ4_XS 37.902661760 GB, Q2_K 26.375108736 GB, Q3_K_L 37.140592768 GB, Q3_K_M 34.267494528 GB, Q3_K_S 30.912051328 GB, Q4_K_M 42.520393856 GB, Q4_K_S 40.347220096 GB, Q5_K_M 49.949816960 GB, Q5_K_S 48.657447040 GB, and split Q6_K total 57.888144192 GB. The selected IQ1_M artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Meta Llama 3.1 70B Instruct IQ1_M GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF/resolve/c1c0c3d88f2fca6f18de9063917f9bc49a33365a/Meta-Llama-3.1-70B-Instruct.IQ1_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 33 metadata entries and 723 tensors. The linked file is 16.751196288 GB. Tensor spans sum to 16.743333888 GB: token_embd.weight 0.344752128 GB, blk.* tensors 15.676211200 GB, output_norm.weight 0.000032768 GB, and output.weight 0.722337792 GB. Metadata/tokenizer/header/file overhead accounts for 0.007862400 GB. Stored tensor bytes split into IQ1_M 13.138657280 GB, IQ2_XXS 1.384120320 GB, Q2_K 1.115455488 GB, Q4_K 0.377487360 GB, Q5_K 0.722337792 GB, and F32 0.005275648 GB. The header records general.architecture llama, Llama 3.1 license, llama.block_count 80, context_length 131072, embedding_length 8192, feed_forward_length 28672, attention.head_count 64, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 500000, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Meta Llama 3.1 70B Instruct GGUF missing config check", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF/raw/c1c0c3d88f2fca6f18de9063917f9bc49a33365a/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json returns 404, and the API config object is empty. This profile intentionally uses the selected GGUF header as the architecture source of truth." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, gated base API/access checks, audited Hugging Quants AWQ config comparison, expanded linked-file size metadata for all GGUF siblings, missing repo-config check, and a direct GGUF header/tensor-index range read of the selected IQ1_M artifact." }, "notes": "Use this profile for the API-selected Meta Llama 3.1 70B Instruct IQ1_M GGUF artifact. Do not infer the lower-bit or split-file sibling footprints from this profile." }, { "id": "maziyarpanahi--meta-llama-3-1-8b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF", "title": "MaziyarPanahi Meta Llama 3.1 8B Instruct GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Meta Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, gated Meta base-profile checks, and selected GGUF header metadata", "config_compatible": false, "notes": "The GGUF repo metadata identifies meta-llama/Llama-3.1-8B-Instruct as the quantized base. The Meta base profile is gated in this audit environment, so this profile uses the selected public GGUF header as direct architecture evidence instead of copying the gated base config." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 3.179131808, "swept_weight_gb": 2.998919168, "auxiliary_resident_weight_gb": 0.18021264, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Meta-Llama-3.1-8B-Instruct.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode excludes token_embd.weight input lookup and includes blk.* tensors, output.weight, and output_norm.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file alignment are resident in the selected artifact but not swept as full matrices for each generated token", "notes": "The selected Q2_K linked file is 3.179131808 GB. Header tensor spans total 3.171295232 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007836576 GB. The main GGUF stores token_embd.weight separately from output.weight, so ordinary text decode excludes the input embedding lookup and charges the separate output projection." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 Llama blocks, 32 attention heads, 8 KV heads, 4096 hidden size, and 128 RoPE/head dimension. No sliding-window setting is present, so this profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. Smaller or larger Q/I-quant artifacts in the repo require separate selected-artifact profiles." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3958939453920989, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor payload bytes for swept weight traffic. Dequantization, llama.cpp kernels, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Q2_K because gguf.totalFileSize exactly matches Meta-Llama-3.1-8B-Instruct.Q2_K.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "MaziyarPanahi Meta Llama 3.1 8B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 1f301d86d760b435a11a56de3863bc0121bfb98f, the API records a public non-gated text-generation GGUF repo with base_model meta-llama/Llama-3.1-8B-Instruct, llama3.1 license, region:us, imatrix metadata, 136669 downloads, GGUF architecture llama, 131072 context length, gguf.total 8030261248, and gguf.totalFileSize 3179131808. The API totalFileSize matches Meta-Llama-3.1-8B-Instruct.Q2_K.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Meta Llama 3.1 8B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF/raw/1f301d86d760b435a11a56de3863bc0121bfb98f/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "selected_artifact", "license", "runtime_format" ], "notes": "The card records base_model meta-llama/Meta-Llama-3.1-8B-Instruct, quantized_by MaziyarPanahi, license_name llama3.1, and GGUF files for llama.cpp-compatible runtimes." }, { "label": "Meta Llama 3.1 8B Instruct gated base profile", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "unsupported_gated_base" ], "notes": "The existing local unsupported profile for meta-llama/Llama-3.1-8B-Instruct records API safetensors total 8030261248 parameters, but config and tensor headers are gated. This GGUF derivative is audited from its public selected artifact instead of inferring from the gated base profile." }, { "label": "MaziyarPanahi Meta Llama 3.1 8B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF/tree/1f301d86d760b435a11a56de3863bc0121bfb98f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 3.179131808 GB, Q3_K_M 4.018918304 GB, Q4_K_M 4.920734624 GB, Q5_K_M 5.732987808 GB, and Q8_0 8.540771232 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Meta Llama 3.1 8B Instruct Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/1f301d86d760b435a11a56de3863bc0121bfb98f/Meta-Llama-3.1-8B-Instruct.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 33 metadata entries and 291 tensors. The linked file is 3.179131808 GB. Tensor spans sum to 3.171295232 GB: token_embd.weight 0.172376064 GB, output.weight 0.430940160 GB, blk.* tensors 2.567962624 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.007836576 GB. Stored tensor spans split into Q2_K 1.625702400 GB, Q3_K 1.038090240 GB, Q4_K 0.075497472 GB, Q6_K 0.430940160 GB, and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 131072, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, RoPE dimension/head dimension 128, RoPE base 500000, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Meta Llama 3.1 8B Instruct GGUF missing repo config check", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF/raw/1f301d86d760b435a11a56de3863bc0121bfb98f/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A pinned repo-local config.json request returned 404. The profile therefore uses the selected GGUF header directly and does not claim repo-local config evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, existing gated-base profile evidence, linked GGUF file sizes, pinned missing repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected Meta Llama 3.1 8B Instruct Q2_K GGUF artifact. Do not silently substitute Q4_K_M, Q8_0, or other artifacts; those require separate profiles with their own selected artifact bytes." }, { "id": "maziyarpanahi--meta-llama-3-8b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF", "title": "MaziyarPanahi Meta Llama 3 8B Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Meta Llama 3 8B Instruct.", "model_family": "llama3-dense-gguf", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3-8B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, original-card text bundled in the GGUF card, and public selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of meta-llama/Meta-Llama-3-8B-Instruct. The base repo remains gated in this audit environment, and the package config.json incorrectly records only model_type mistral, so this profile uses the selected public GGUF header as the architecture source instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "llama-3-8b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 16.06940384, "swept_weight_gb": 15.010381824, "auxiliary_resident_weight_gb": 1.059022016, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Meta-Llama-3-8B-Instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices for each generated token", "notes": "The selected FP16 linked file is 16.069403840 GB. GGUF tensor spans total 16.061054976 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.008348864 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 decoder layers, 8 KV heads, 128-dimensional key/value heads, and 8192 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The repo contains many lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches Meta-Llama-3-8B-Instruct.fp16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant FP16 tensor format." }, "evidence": [ { "label": "MaziyarPanahi Meta Llama 3 8B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 4ece958b356f2ec00338e5898ba0a7254d976baf, the live API records a public non-gated GGUF text-generation repo with base_model meta-llama/Meta-Llama-3-8B-Instruct, base_model:quantized metadata, region:us, 142135 downloads, GGUF architecture llama, context_length 8192, gguf.total 8030261248, and gguf.totalFileSize 16069403840. The API totalFileSize matches Meta-Llama-3-8B-Instruct.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Meta Llama 3 8B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/raw/4ece958b356f2ec00338e5898ba0a7254d976baf/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "architecture_compatibility" ], "notes": "The card records base_model meta-llama/Meta-Llama-3-8B-Instruct, quantized_by MaziyarPanahi, license_name llama3, and GGUF files for llama.cpp-compatible runtimes. The bundled original README states that Llama 3 8B uses an optimized autoregressive transformer architecture, 8k context, and Grouped-Query Attention. The card advertises multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a smaller default." }, { "label": "MaziyarPanahi Meta Llama 3 8B Instruct GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/raw/4ece958b356f2ec00338e5898ba0a7254d976baf/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral, which conflicts with the live API GGUF architecture and the selected GGUF header. This profile therefore does not use that config as architecture evidence." }, { "label": "Meta Llama 3 8B Instruct gated base profile", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/raw/8afb486c1db24fe5011ec46dfbe5b5dccdb575c2/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The existing base profile records that raw config, README, and safetensors index requests returned access denied in this audit environment. This GGUF profile therefore does not claim a direct base-config comparison, but it can use the public selected GGUF header for the served GGUF artifact's exact geometry." }, { "label": "MaziyarPanahi Meta Llama 3 8B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/tree/4ece958b356f2ec00338e5898ba0a7254d976baf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 16.069403840 GB, Q8_0 8.541283552 GB, Q4_K_M 4.921246944 GB, Q2_K 3.179644128 GB, and IQ1_M 2.162484448 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Meta Llama 3 8B Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/resolve/4ece958b356f2ec00338e5898ba0a7254d976baf/Meta-Llama-3-8B-Instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 21 metadata entries and 291 tensors. The linked file is 16.069403840 GB. Tensor spans sum to 16.061054976 GB: token_embd.weight 1.050673152 GB, output.weight 1.050673152 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.008348864 GB. Stored tensor spans split into F16 16.059990016 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 8192, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 500000, vocab_size 128256, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card metadata, package config conflict check, existing gated-base access profile, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Meta Llama 3 8B Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--ministral-3-3b-reasoning-2512-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF", "title": "MaziyarPanahi Ministral 3 3B Reasoning 2512 GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Ministral 3 3B Reasoning 2512.", "model_family": "ministral3-dense-gguf", "base_model_proof": { "base_model": "mistralai/Ministral-3-3B-Reasoning-2512", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object size checks, and Mistral3 base text_config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mistralai/Ministral-3-3B-Reasoning-2512. The base and base-base configs record Mistral3ForConditionalGeneration with a Ministral3 text_config and Pixtral vision_config. The selected GGUF header records the same text-tower geometry and no vision tensors, so compatibility is true for ordinary text decode only." }, "architecture": { "canonical_architecture_id": "ministral-3-3b-reasoning-2512-text", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.429006336, "swept_params_b": 3.429006336, "auxiliary_resident_params_b": 0, "resident_weight_gb": 6.8662192, "swept_weight_gb": 6.858338304, "auxiliary_resident_weight_gb": 0.007880896, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Ministral-3-3B-Reasoning-2512.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors", "notes": "The selected FP16 linked file is 6.866219200 GB. GGUF tensor spans total 6.858338304 GB, while metadata, tokenizer, header, and file overhead account for 0.007880896 GB. The selected text GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight, but no output.weight tensor; token_embd.weight is therefore charged as tied output projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 26, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header and base text_config record 26 text decoder layers, 8 KV heads, 128-dimensional key/value heads, sliding_window null, and 262144 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected text GGUF artifact after any prompt processing. It does not include the base model's Pixtral vision tower or multimodal projector because the selected GGUF header contains only text tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, multimodal prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Ministral-3-3B-Reasoning-2512.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Ministral 3 3B Reasoning GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75, the API records a public non-gated GGUF text-generation repo with base_model mistralai/Ministral-3-3B-Reasoning-2512, quantized tags, region:us, 113921 downloads, GGUF architecture mistral3, context length 262144, gguf.total 3429006336, and gguf.totalFileSize 6866219200. The API totalFileSize matches Ministral-3-3B-Reasoning-2512.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Ministral 3 3B Reasoning GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/raw/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model mistralai/Ministral-3-3B-Reasoning-2512, model_creator mistralai, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists multiple lower-bit quantized siblings, but the API-selected totalFileSize points at the FP16 sibling." }, { "label": "MaziyarPanahi Ministral GGUF package config stub", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/raw/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75/config.json", "source_type": "config", "supports": [ "package_config_scope" ], "notes": "The package config is only 31 bytes and contains {\"model_type\":\"mistral\"}. The selected GGUF header and base Mistral3 configs provide the actual architecture evidence, so this stub is not used for bounds geometry." }, { "label": "Mistral AI Ministral 3 3B Reasoning API metadata", "url": "https://huggingface.co/api/models/mistralai/Ministral-3-3B-Reasoning-2512", "source_type": "model_card", "supports": [ "base_model_proof", "license", "base_lineage" ], "notes": "At commit 942382969cbb4a5352d59550c3373dee88e8bf6a, the base repo is public, Apache-2.0 licensed, and records base_model mistralai/Ministral-3-3B-Base-2512 plus safetensors total 4251743232 parameters. The base total includes non-text components, while this GGUF profile targets the selected text-only artifact." }, { "label": "Mistral AI Ministral 3 3B Base config", "url": "https://huggingface.co/mistralai/Ministral-3-3B-Base-2512/raw/6f9c4b12a95b139af68670a6713616b757923735/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The base config records Mistral3ForConditionalGeneration, dtype bfloat16, text_config model_type ministral3, 26 text layers, hidden size 3072, intermediate size 9216, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max positions, sliding_window null, tie_word_embeddings true, and vocab_size 131072. It also records a Pixtral vision_config, which is outside this text GGUF profile." }, { "label": "MaziyarPanahi Ministral GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/tree/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata and HEAD checks found Q2_K 1.458958528 GB, Q3_K_M 1.795551424 GB, Q3_K_L 1.934356672 GB, Q4_K_M 2.146496704 GB, Q5_K_M 2.473652416 GB, Q6_K 2.821255360 GB, and FP16 6.866219200 GB. The FP16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Ministral FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/resolve/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75/Ministral-3-3B-Reasoning-2512.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 50 metadata entries and 236 tensors. The selected file is 6.866219200 GB, with tensor payloads starting at byte 7880896. Tensor spans sum to 6.858338304 GB across 3.429006336B logical elements: token_embd.weight 0.805306368 GB, blk.* tensors 6.053019648 GB, and output_norm.weight 0.000012288 GB. Metadata/tokenizer/header/file overhead accounts for 0.007880896 GB. Stored tensor bytes split into F16 6.857687040 GB and F32 0.000651264 GB. The header records general.architecture mistral3, Apache-2.0 license, block_count 26, context_length 262144, embedding_length 3072, feed_forward_length 9216, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope scaling yarn factor 16, original context 16384, and no separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, package config-stub check, base/base-base API and config metadata, selected linked-object size checks, and direct selected FP16 GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected Ministral 3 3B Reasoning FP16 text GGUF artifact. Do not silently substitute the lower-bit siblings, the base BF16 multimodal checkpoint, or the Pixtral vision tower." }, { "id": "maziyarpanahi--mistral-7b-instruct-v0-3-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF", "title": "MaziyarPanahi Mistral 7B Instruct v0.3 GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Mistral 7B Instruct v0.3.", "model_family": "mistral-7b-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-7B-Instruct-v0.3", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and pinned base config comparison", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of mistralai/Mistral-7B-Instruct-v0.3. The selected FP16 GGUF header records the same Mistral text geometry as the pinned base config: 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 key/value head dimension, 32768 context, rope theta 1000000, and untied input/output embeddings." }, "architecture": { "canonical_architecture_id": "mistral-7b-v0-3", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.248023552, "swept_params_b": 7.113805824, "auxiliary_resident_params_b": 0.134217728, "resident_weight_gb": 14.497337312, "swept_weight_gb": 14.228144128, "auxiliary_resident_weight_gb": 0.269193184, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Mistral-7B-Instruct-v0.3.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 14.497337312 GB. GGUF tensor spans total 14.496579584 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.000757728 GB. Because output.weight is stored separately and the base config records tie_word_embeddings false, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header and pinned base config record 32 decoder layers, 8 KV heads, 128-dimensional key/value heads, sliding_window null, and 32768 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The repo contains many lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF because gguf.totalFileSize exactly matches Mistral-7B-Instruct-v0.3.fp16.gguf. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral 7B Instruct v0.3 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit ce89f595755a4bf2e2e05d155cc43cb847c78978, the API records a public non-gated Apache-2.0 GGUF text-generation repo with base_model mistralai/Mistral-7B-Instruct-v0.3, base_model:quantized metadata, endpoints_compatible, region:us, 142761 downloads, GGUF architecture llama, context_length 32768, gguf.total 7248023552, and gguf.totalFileSize 14497337312. The API totalFileSize matches Mistral-7B-Instruct-v0.3.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral 7B Instruct v0.3 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format" ], "notes": "The card records Apache-2.0 licensing, original model mistralai/Mistral-7B-Instruct-v0.3, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. It tags multiple lower-bit quantizations but does not override the API-selected FP16 artifact with a normal-serving default." }, { "label": "Mistral 7B Instruct v0.3 pinned base config", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/raw/c170c708c41dac9275d15a8fff4eca08d52bab71/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility", "embedding_layout", "kv_adapter" ], "notes": "The pinned base config records MistralForCausalLM, bfloat16, 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 32768 max positions, sliding_window null, tie_word_embeddings false, vocab_size 32768, and rope_theta 1000000." }, { "label": "MaziyarPanahi Mistral 7B Instruct v0.3 GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF/tree/ce89f595755a4bf2e2e05d155cc43cb847c78978", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 14.497337312 GB, Q8_0 7.702565024 GB, Q4_K_M 4.372811936 GB, Q2_K 2.722877600 GB, and IQ1_M 1.757663392 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mistral 7B Instruct v0.3 FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF/resolve/ce89f595755a4bf2e2e05d155cc43cb847c78978/Mistral-7B-Instruct-v0.3.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 25 metadata entries and 291 tensors. The linked file is 14.497337312 GB. Tensor spans sum to 14.496579584 GB: output.weight 0.268435456 GB, token_embd.weight 0.268435456 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.000757728 GB. Stored tensor spans split into F16 14.495514624 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 32768, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, and rope.freq_base 1000000." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned base config, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Mistral 7B Instruct v0.3 FP16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "maziyarpanahi--mistral-large-instruct-2411-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF", "title": "MaziyarPanahi Mistral Large Instruct 2411 GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Mistral Large Instruct 2411.", "model_family": "mistral-large-2411-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-Large-Instruct-2411", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, package-config absence check, and pinned Mistral base config", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of mistralai/Mistral-Large-Instruct-2411. The pinned base config records MistralForCausalLM with 88 layers, hidden size 12288, intermediate size 28672, 96 attention heads, 8 KV heads, 131072 max positions, sliding_window null, and untied embeddings. The selected GGUF header uses the llama.cpp llama architecture label but records matching checked geometry and context fields." }, "architecture": { "canonical_architecture_id": "mistral-large-2411", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 122.610069504, "swept_params_b": 122.20741632, "auxiliary_resident_params_b": 0.402653184, "resident_weight_gb": 45.196298272, "swept_weight_gb": 45.063389184, "auxiliary_resident_weight_gb": 0.132909088, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Mistral-Large-Instruct-2411.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 45.196298272 GB. GGUF tensor spans total 45.195509760 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000788512 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 88, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header and base config record 88 decoder blocks, 8 KV heads, 128-dimensional key/value heads, 131072 context, and no sliding-window cap. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3686181604401222, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, kernels, scheduler behavior, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets Mistral-Large-Instruct-2411.Q2_K.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral Large Instruct 2411 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 76081fc502213f8e894d6089bc648fa4fcc962c9, the API records a public non-gated GGUF text-generation repo with base_model mistralai/Mistral-Large-Instruct-2411, base_model:quantized metadata, quantized tags, region:us, 111457 downloads, GGUF architecture llama, context_length 131072, gguf.total 122610069504, and gguf.totalFileSize 45196298272. The API totalFileSize matches Mistral-Large-Instruct-2411.Q2_K.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral Large Instruct 2411 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF/raw/76081fc502213f8e894d6089bc648fa4fcc962c9/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model mistralai/Mistral-Large-Instruct-2411, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF header and base model record Mistral Research License metadata." }, { "label": "Mistral Large Instruct 2411 base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mistral-Large-Instruct-2411", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit ba78820945ae22361b0274cf0ae6d696c967c1a4, the base repo is public, non-gated, vLLM tagged, region:us tagged, and records BF16 safetensors total 122610069504 parameters. The card metadata records license other, license_name mrl, and a Mistral AI Research License link." }, { "label": "Mistral Large Instruct 2411 base config", "url": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411/raw/ba78820945ae22361b0274cf0ae6d696c967c1a4/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The base config records MistralForCausalLM, 88 layers, hidden size 12288, intermediate size 28672, 96 attention heads, 8 KV heads, 128 head_dim, 131072 max positions, sliding_window null, rope_theta 1000000, use_cache true, tie_word_embeddings false, and vocab size 32768." }, { "label": "MaziyarPanahi Mistral Large GGUF selected linked-object check", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF/resolve/76081fc502213f8e894d6089bc648fa4fcc962c9/Mistral-Large-Instruct-2411.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF linked-object metadata reports Mistral-Large-Instruct-2411.Q2_K.gguf size 45.196298272 GB, matching API gguf.totalFileSize and selecting the only single-file GGUF artifact in the repo. Other quantization families are split across seven GGUF part files and are not selected by the API totalFileSize." }, { "label": "MaziyarPanahi Mistral Large Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF/resolve/76081fc502213f8e894d6089bc648fa4fcc962c9/Mistral-Large-Instruct-2411.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "license" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 37 metadata entries and 795 tensors. The linked file is 45.196298272 GB, with tensor data beginning at byte 788512. Tensor spans sum to 45.195509760 GB: output.weight 0.330301440 GB, output_norm.weight 0.000049152 GB, token_embd.weight 0.132120576 GB, and blk.* tensors 44.733038592 GB. Metadata/tokenizer/header/file overhead accounts for 0.000788512 GB. Stored tensor bytes split into Q2_K 25.201999872 GB, Q3_K 19.031654400 GB, Q4_K 0.622854144 GB, Q6_K 0.330301440 GB, and F32 0.008699904 GB. The header records general.architecture llama, license other, license_name mrl, Mistral AI Research License link, 88 blocks, 131072 context, 12288 embedding length, 28672 feed-forward length, 96 attention heads, 8 KV heads, key/value length 128, rope.freq_base 1000000, vocab size 32768, and separate output.weight." }, { "label": "MaziyarPanahi Mistral Large GGUF package config absence", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF/raw/76081fc502213f8e894d6089bc648fa4fcc962c9/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package has no repo-local config.json; the raw config request returned HTTP 404. This profile therefore uses the selected GGUF header plus the pinned Mistral base config as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, selected linked-file metadata, package config absence check, and a direct GGUF header/tensor-index range read of the selected Q2_K artifact." }, "notes": "Use this profile for the API-selected Mistral Large Instruct 2411 Q2_K GGUF artifact. Do not infer footprints for the split GGUF sibling quantizations from this profile." }, { "id": "maziyarpanahi--mistral-nemo-instruct-2407-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF", "title": "MaziyarPanahi Mistral Nemo Instruct 2407 GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Mistral Nemo Instruct 2407.", "model_family": "mistral-nemo-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-Nemo-Instruct-2407", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object HEAD checks, and Mistral Nemo base config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mistralai/Mistral-Nemo-Instruct-2407. The base config records MistralForCausalLM geometry with 40 layers, 32 attention heads, 8 KV heads, 128 head dimension, 5120 hidden size, untied embeddings, no sliding window, and 131072 max positions. The selected GGUF header records llama.cpp architecture llama with the same Mistral-Nemo-compatible layer, attention, KV, head-dimension, RoPE, and untied embedding geometry, but a 1024000-token context field." }, "architecture": { "canonical_architecture_id": "mistral-nemo-instruct-2407", "max_context_tokens": 1024000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 12.2477824, "swept_params_b": 11.57669376, "auxiliary_resident_params_b": 0.67108864, "resident_weight_gb": 24.504276704, "swept_weight_gb": 23.15421696, "auxiliary_resident_weight_gb": 1.350059744, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Mistral-Nemo-Instruct-2407.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 24.504276704 GB. GGUF tensor payloads total 24.496394240 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007882464 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 1024000 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Mistral-Nemo-Instruct-2407.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral Nemo GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit eba4e7492de28b8ab2ff44b0bb819004181b3db4, the API records a public non-gated GGUF repo with base_model mistralai/Mistral-Nemo-Instruct-2407, quantized tags, region:us, text-generation pipeline, 123676 downloads, GGUF architecture llama, context length 1024000, gguf.total 12247782400, and gguf.totalFileSize 24504276704. The API totalFileSize matches Mistral-Nemo-Instruct-2407.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral Nemo GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF/raw/eba4e7492de28b8ab2ff44b0bb819004181b3db4/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The model card records base_model mistralai/Mistral-Nemo-Instruct-2407, model_creator mistralai, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists many quantized siblings, but the API-selected totalFileSize points at the F16 sibling." }, { "label": "Mistral Nemo Instruct base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mistral-Nemo-Instruct-2407", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit 04d8a90549d23fc6bd7f642064003592df51e9b3, the base repo is public, Apache-2.0 licensed, and records BF16 safetensors parameters 12247782400." }, { "label": "Mistral Nemo Instruct base config", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407/raw/04d8a90549d23fc6bd7f642064003592df51e9b3/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, 40 layers, hidden size 5120, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 131072, sliding_window null, rope_theta 1000000, and tie_word_embeddings false." }, { "label": "MaziyarPanahi Mistral Nemo GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF/tree/eba4e7492de28b8ab2ff44b0bb819004181b3db4", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 4.791048128 GB, Q3_K_L 6.561503168 GB, Q3_K_M 6.083090368 GB, Q3_K_S 5.534226368 GB, Q4_K_M 7.477204928 GB, Q4_K_S 7.120197568 GB, Q5_K_M 8.727631808 GB, Q5_K_S 8.518735808 GB, Q6_K 10.056210368 GB, Q8_0 13.022369728 GB, and F16 24.504276704 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mistral Nemo F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF/resolve/eba4e7492de28b8ab2ff44b0bb819004181b3db4/Mistral-Nemo-Instruct-2407.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 363 tensors. The selected file is 24.504276704 GB, with tensor payloads starting at byte 7882464. Tensor payloads sum to 24.496394240 GB across 12.247782400B logical elements: token_embd.weight 1.342177280 GB, output.weight 1.342177280 GB, blk.* tensors 21.812019200 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.007882464 GB. Stored tensor bytes split into F16 24.494735360 GB and F32 0.001658880 GB. The header records general.architecture llama, llama.block_count 40, context_length 1024000, embedding_length 5120, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000, and vocab_size 131072." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, base Mistral Nemo API/config metadata, selected linked-object HEAD checks, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--mistral-small-24b-instruct-2501-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF", "title": "MaziyarPanahi Mistral Small 24B Instruct 2501 GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Mistral Small 24B Instruct 2501.", "model_family": "mistral-small-24b-2501-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-Small-24B-Instruct-2501", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object HEAD checks, and Mistral Small base config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mistralai/Mistral-Small-24B-Instruct-2501. The base config records MistralForCausalLM geometry with 40 layers, 32 attention heads, 8 KV heads, 128 head dimension, 5120 hidden size, untied embeddings, no sliding window, and 32768 max positions. The selected GGUF header records llama.cpp architecture llama with the same Mistral-compatible layer, attention, KV, head-dimension, RoPE, context, and untied embedding geometry." }, "architecture": { "canonical_architecture_id": "mistral-small-24b-2501", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 23.5724032, "swept_params_b": 22.90131456, "auxiliary_resident_params_b": 0.67108864, "resident_weight_gb": 47.153518016, "swept_weight_gb": 45.80345856, "auxiliary_resident_weight_gb": 1.350059456, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Mistral-Small-24B-Instruct-2501.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 47.153518016 GB. GGUF tensor payloads total 47.145635840 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007882176 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 32768 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Mistral-Small-24B-Instruct-2501.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral Small 24B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 458f84af662e5b23b03e8191e9c3c4d6ab857283, the API records a public non-gated GGUF repo with base_model mistralai/Mistral-Small-24B-Instruct-2501, quantized tags, region:us, text-generation pipeline, Apache-2.0 license, 116025 downloads, GGUF architecture llama, context length 32768, gguf.total 23572403200, and gguf.totalFileSize 47153518016. The API totalFileSize matches Mistral-Small-24B-Instruct-2501.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral Small 24B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/raw/458f84af662e5b23b03e8191e9c3c4d6ab857283/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "selected_artifact" ], "notes": "The model card records base_model mistralai/Mistral-Small-24B-Instruct-2501, model_creator mistralai, quantized_by MaziyarPanahi, Apache-2.0 license, and GGUF files for llama.cpp-compatible runtimes. It lists many quantized siblings, but the API-selected totalFileSize points at the F16 sibling." }, { "label": "Mistral Small 24B Instruct base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mistral-Small-24B-Instruct-2501", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit 9527884be6e5616bdd54de542f9ae13384489724, the base repo is public, Apache-2.0 licensed, and records BF16 safetensors parameters 23572403200." }, { "label": "Mistral Small 24B Instruct base config", "url": "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501/raw/9527884be6e5616bdd54de542f9ae13384489724/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, 40 layers, hidden size 5120, intermediate size 32768, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, sliding_window null, rope_theta 100000000, tie_word_embeddings false, and vocab_size 131072." }, { "label": "MaziyarPanahi Mistral Small 24B GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/raw/458f84af662e5b23b03e8191e9c3c4d6ab857283/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral. That is consistent with the base family but insufficient for geometry, so this profile uses the selected GGUF header and public base config for architecture fields." }, { "label": "MaziyarPanahi Mistral Small 24B GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/tree/458f84af662e5b23b03e8191e9c3c4d6ab857283", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 8.890324416 GB, Q3_K_L 12.400760256 GB, Q3_K_M 11.474081216 GB, Q3_K_S 10.400273856 GB, Q4_K_M 14.333908416 GB, Q4_K_S 13.549278656 GB, Q5_K_M 16.763983296 GB, Q5_K_S 16.304412096 GB, Q6_K 19.345937856 GB, Q8_0 25.054778816 GB, and F16 47.153518016 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mistral Small 24B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/resolve/458f84af662e5b23b03e8191e9c3c4d6ab857283/Mistral-Small-24B-Instruct-2501.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 40 metadata entries and 363 tensors. The selected file is 47.153518016 GB, with tensor payloads starting at byte 7882176. Tensor payloads sum to 47.145635840 GB across 23.572403200B logical elements: token_embd.weight 1.342177280 GB, output.weight 1.342177280 GB, blk.* tensors 44.461260800 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.007882176 GB. Stored tensor bytes split into F16 47.143976960 GB and F32 0.001658880 GB. The header records general.architecture llama, llama.block_count 40, context_length 32768, embedding_length 5120, feed_forward_length 32768, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 100000000, and vocab_size 131072." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, base Mistral Small API/config metadata, selected linked-object HEAD checks, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--mistral-small-3-1-24b-instruct-2503-hf-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF", "title": "MaziyarPanahi Mistral Small 3.1 24B Instruct 2503 HF GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Mistral Small 3.1 24B Instruct 2503 HF.", "model_family": "mistral-small-3-1-24b-2503-dense-gguf", "base_model_proof": { "base_model": "mrfakename/mistral-small-3.1-24b-instruct-2503-hf", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object size checks, and Mistral Small 3.1 base-wrapper config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mrfakename/mistral-small-3.1-24b-instruct-2503-hf, whose base wrapper records a finetune relationship to mistralai/Mistral-Small-3.1-24B-Instruct-2503. The base-wrapper config records MistralForCausalLM geometry with 40 layers, 32 attention heads, 8 KV heads, 128 head dimension, 5120 hidden size, untied embeddings, no sliding window, RoPE theta 1000000000, and 32768 max positions. The selected GGUF header records llama.cpp architecture llama with the same Mistral-compatible layer, attention, KV, head-dimension, RoPE, context, and untied embedding geometry." }, "architecture": { "canonical_architecture_id": "mistral-small-3-1-24b-2503-hf", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 23.5724032, "swept_params_b": 22.90131456, "auxiliary_resident_params_b": 0.67108864, "resident_weight_gb": 47.153519136, "swept_weight_gb": 45.80345856, "auxiliary_resident_weight_gb": 1.350060576, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for mistral-small-3.1-24b-instruct-2503-hf.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 47.153519136 GB. GGUF tensor payloads total 47.145635840 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007883296 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 32768 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is mistral-small-3.1-24b-instruct-2503-hf.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral Small 3.1 24B HF GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 69f1da83040e31989aa85ae485e3dff329d954ee, the API records a public non-gated GGUF repo with base_model mrfakename/mistral-small-3.1-24b-instruct-2503-hf, quantized tags, region:us, text-generation pipeline, 113869 downloads, GGUF architecture llama, context length 32768, gguf.total 23572403200, and gguf.totalFileSize 47153519136. The API totalFileSize matches mistral-small-3.1-24b-instruct-2503-hf.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral Small 3.1 24B HF GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF/raw/69f1da83040e31989aa85ae485e3dff329d954ee/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The model card records base_model mrfakename/mistral-small-3.1-24b-instruct-2503-hf, model_creator mrfakename, quantized_by MaziyarPanahi, and GGUF files for llama.cpp-compatible runtimes. It lists many quantized siblings, but the API-selected totalFileSize points at the F16 sibling." }, { "label": "Mistral Small 3.1 24B HF base-wrapper API metadata", "url": "https://huggingface.co/api/models/mrfakename/mistral-small-3.1-24b-instruct-2503-hf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit 822b1dd72c3f51176c7519290f1882593d5a10b0, the base-wrapper repo is public, Apache-2.0 licensed, references mistralai/Mistral-Small-3.1-24B-Instruct-2503 as base_model, and records BF16 safetensors parameters 23572403200." }, { "label": "Mistral Small 3.1 24B HF base-wrapper config", "url": "https://huggingface.co/mrfakename/mistral-small-3.1-24b-instruct-2503-hf/raw/822b1dd72c3f51176c7519290f1882593d5a10b0/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, 40 layers, hidden size 5120, intermediate size 32768, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, sliding_window null, rope_theta 1000000000, tie_word_embeddings false, and vocab_size 131072." }, { "label": "MaziyarPanahi Mistral Small 3.1 HF GGUF package config", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF/raw/69f1da83040e31989aa85ae485e3dff329d954ee/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package config only contains model_type mistral. That is consistent with the base family but insufficient for geometry, so this profile uses the selected GGUF header and public base-wrapper config for architecture fields." }, { "label": "MaziyarPanahi Mistral Small 3.1 HF GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF/tree/69f1da83040e31989aa85ae485e3dff329d954ee", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata found GGUF siblings Q2_K 8.890325536 GB, Q3_K_L 12.400761376 GB, Q3_K_M 11.474082336 GB, Q3_K_S 10.400274976 GB, Q4_K_M 14.333909536 GB, Q4_K_S 13.549279776 GB, Q5_K_M 16.763984416 GB, Q5_K_S 16.304413216 GB, Q6_K 19.345938976 GB, Q8_0 25.054779936 GB, and F16 47.153519136 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mistral Small 3.1 HF F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF/resolve/69f1da83040e31989aa85ae485e3dff329d954ee/mistral-small-3.1-24b-instruct-2503-hf.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 40 metadata entries and 363 tensors. The selected file is 47.153519136 GB, with tensor payloads starting at byte 7883296. Tensor payloads sum to 47.145635840 GB across 23.572403200B logical elements: token_embd.weight 1.342177280 GB, output.weight 1.342177280 GB, blk.* tensors 44.461260800 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.007883296 GB. Stored tensor bytes split into F16 47.143976960 GB and F32 0.001658880 GB. The header records general.architecture llama, Apache-2.0 license metadata, base_model repo URL for mistralai/Mistral-Small-3.1-24B-Instruct-2503, llama.block_count 40, context_length 32768, embedding_length 5120, feed_forward_length 32768, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000000, vocab_size 131072, and no llama.attention.sliding_window field." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, base-wrapper Mistral Small 3.1 API/config metadata, selected linked-object size checks, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--mistral-small-instruct-2409-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF", "title": "MaziyarPanahi Mistral Small Instruct 2409 GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Mistral Small Instruct 2409.", "model_family": "mistral-small-2409-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-Small-Instruct-2409", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, selected linked-object size checks, and Mistral Small base config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of mistralai/Mistral-Small-Instruct-2409. The base config records MistralForCausalLM geometry with 56 layers, 48 attention heads, 8 KV heads, 128 head dimension, 6144 hidden size, 16384 intermediate size, untied embeddings, no sliding window, and 32768 max positions. The selected GGUF header records llama.cpp architecture llama with matching Mistral-compatible layer, attention, KV, head-dimension, RoPE, context, vocabulary, and untied embedding geometry. The repo-local package config is only a stale model_type mistral stub and is not used as the architecture source." }, "architecture": { "canonical_architecture_id": "mistral-small-instruct-2409", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 22.247282688, "swept_params_b": 22.045956096, "auxiliary_resident_params_b": 0.201326592, "resident_weight_gb": 44.496728832, "swept_weight_gb": 44.093300736, "auxiliary_resident_weight_gb": 0.403428096, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Mistral-Small-Instruct-2409.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 44.496728832 GB. GGUF tensor payloads total 44.495953920 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000774912 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 56, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 56 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 32768 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Mistral-Small-Instruct-2409.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Mistral Small Instruct 2409 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 23bcb4eba278cea53f211652d60dbcd6724f89e2, the API records a public non-gated GGUF repo with base_model mistralai/Mistral-Small-Instruct-2409, quantized tags, region:us, conversational text-generation pipeline, 113311 downloads, GGUF architecture llama, context length 32768, gguf.total 22247282688, and gguf.totalFileSize 44496728832. The API totalFileSize matches Mistral-Small-Instruct-2409.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Mistral Small Instruct 2409 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/raw/23bcb4eba278cea53f211652d60dbcd6724f89e2/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The model card records base_model mistralai/Mistral-Small-Instruct-2409, model_creator mistralai, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists many quantized siblings, but the API-selected totalFileSize points at the F16 sibling." }, { "label": "Mistral Small Instruct 2409 base API metadata", "url": "https://huggingface.co/api/models/mistralai/Mistral-Small-Instruct-2409", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit 4600506f6b13c7ef89e61a54263f4c9bf483de30, the base repo is public, records license:other / Mistral Research License, and records BF16 safetensors parameters 22247282688." }, { "label": "Mistral Small Instruct 2409 base config", "url": "https://huggingface.co/mistralai/Mistral-Small-Instruct-2409/raw/4600506f6b13c7ef89e61a54263f4c9bf483de30/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, 56 layers, hidden size 6144, intermediate size 16384, 48 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 32768, sliding_window null, rope_theta 1000000, vocab size 32768, and tie_word_embeddings false." }, { "label": "Mistral Small Instruct 2409 params.json", "url": "https://huggingface.co/mistralai/Mistral-Small-Instruct-2409/raw/4600506f6b13c7ef89e61a54263f4c9bf483de30/params.json", "source_type": "config", "supports": [ "architecture", "kv_adapter" ], "notes": "The Mistral params.json independently records dim 6144, 56 layers, 48 heads, 8 KV heads, head_dim 128, hidden_dim 16384, vocab_size 32768, max_position_embeddings 32768, and rope_theta 1000000." }, { "label": "MaziyarPanahi Mistral Small Instruct 2409 GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/tree/23bcb4eba278cea53f211652d60dbcd6724f89e2", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ1_S 4.829489280 GB, IQ1_M 5.267138688 GB, IQ2_XS 6.646147200 GB, IQ3_XS 9.176098944 GB, IQ4_XS 11.935295616 GB, Q2_K 8.272098048 GB, Q3_K_S 9.641276160 GB, Q3_K_M 10.756829952 GB, Q3_K_L 11.730432768 GB, Q4_K_S 12.660388608 GB, Q4_K_M 13.341242112 GB, Q5_K_S 15.324820224 GB, Q5_K_M 15.722558208 GB, Q6_K 18.252706560 GB, Q8_0 23.640552192 GB, and fp16 44.496728832 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Mistral Small Instruct 2409 F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/resolve/23bcb4eba278cea53f211652d60dbcd6724f89e2/Mistral-Small-Instruct-2409.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 38 metadata entries and 507 tensors. The selected file is 44.496728832 GB, with tensor payloads starting at byte 774912. Tensor payloads sum to 44.495953920 GB across 22.247282688B logical elements: token_embd.weight 0.402653184 GB, output.weight 0.402653184 GB, blk.* tensors 43.690622976 GB, and output_norm.weight 0.000024576 GB. Metadata/tokenizer/header/file overhead accounts for 0.000774912 GB. Stored tensor bytes split into F16 44.493176832 GB and F32 0.002777088 GB. The header records general.architecture llama, general.license other, general.license.name mrl, llama.block_count 56, context_length 32768, embedding_length 6144, feed_forward_length 16384, attention.head_count 48, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000, vocab_size 32768, and no sliding-window field." }, { "label": "MaziyarPanahi Mistral Small Instruct 2409 GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/raw/23bcb4eba278cea53f211652d60dbcd6724f89e2/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral. It is recorded as a stale package stub and is not used as the main architecture evidence because the selected GGUF header and pinned base config provide the full served geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, base Mistral Small API/config metadata, params.json, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--mixtral-8x22b-v0-1-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF", "title": "MaziyarPanahi Mixtral 8x22B v0.1 GGUF IQ1_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected IQ1_M GGUF artifact of Mixtral 8x22B v0.1.", "model_family": "mixtral-moe", "base_model_proof": { "base_model": "v2ray/Mixtral-8x22B-v0.1", "relation": "quantized", "source": "Hugging Face model card/API metadata, target config, base config, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of v2ray/Mixtral-8x22B-v0.1. The target config and base config match the selected GGUF header geometry: MixtralForCausalLM, 56 layers, 48 attention heads, 8 KV heads, 128 key/value head dimension, 8 local experts, 2 experts per token, untied embeddings, and 65536 context tokens." }, "architecture": { "canonical_architecture_id": "mixtral-8x22b-v0.1", "max_context_tokens": 65536, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.732397728, "main_resident_weight_gb": 32.667131904, "auxiliary_resident_weight_gb": 0.065265824, "fixed_weight_gb": 2.455560192, "routed_expert_weight_gb": 3.776446464, "routed_experts": 8, "routed_experts_per_token": 2, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "API-selected Mixtral-8x22B-v0.1.IQ1_M.gguf linked file size including GGUF metadata, tokenizer, header, alignment, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.55 non-routed tensors, routers, and expected-distinct routed expert tensor groups from the selected IQ1_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "Header-derived stored bytes are used instead of rounded 141B/35B model-card parameters. The selected main GGUF mixes IQ1_M, Q2_K, Q4_K, Q5_K, F16, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 8 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 56, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata, target config, and base config record 56 full-context attention layers with 8 KV heads and 128 key/value dimensions. sliding_window is null. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected main IQ1_M GGUF artifact. The repo contains larger split Q2_K, Q3, Q4, Q5, Q6, Q8_0, and FP16 GGUF siblings that require separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.23277094385685712, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq1-m-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and any runtime-specific expert routing locality are outside Bounds Engine v1.", "notes": "The API-selected artifact is Mixtral-8x22B-v0.1.IQ1_M.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "MaziyarPanahi Mixtral 8x22B v0.1 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 053abe826eb9c5b1794e97bba22f0c50910e54e5, the API records a public non-gated Apache-2.0 GGUF repo with base_model v2ray/Mixtral-8x22B-v0.1, base_model:quantized, region:us, 131580 downloads, GGUF architecture llama, context length 65536, gguf.total 140620634112, and gguf.totalFileSize 32732397728. The totalFileSize matches the IQ1_M linked object." }, { "label": "MaziyarPanahi Mixtral 8x22B v0.1 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/raw/053abe826eb9c5b1794e97bba22f0c50910e54e5/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format" ], "notes": "The card records Apache-2.0 licensing, base_model v2ray/Mixtral-8x22B-v0.1, quantized_by MaziyarPanahi, standard GGUF/llama.cpp runtime guidance, and a Mixtral 8x22B description with roughly 141B total and 35B active parameters." }, { "label": "MaziyarPanahi Mixtral 8x22B v0.1 GGUF config", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/raw/053abe826eb9c5b1794e97bba22f0c50910e54e5/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The repo config records MixtralForCausalLM, model_type mixtral, bfloat16 source dtype, 56 layers, hidden_size 6144, intermediate_size 16384, 48 attention heads, 8 KV heads, 8 local experts, 2 experts per token, tie_word_embeddings false, sliding_window null, rope_theta 1000000, and 65536 max position embeddings." }, { "label": "v2ray Mixtral 8x22B v0.1 base config", "url": "https://huggingface.co/v2ray/Mixtral-8x22B-v0.1/raw/701e2d7a02ab9fea587f2a71bda2b98d725adb3b/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The base config at commit 701e2d7a02ab9fea587f2a71bda2b98d725adb3b has the same audited geometry fields as the GGUF repo config and selected GGUF metadata." }, { "label": "MaziyarPanahi Mixtral 8x22B v0.1 GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/tree/053abe826eb9c5b1794e97bba22f0c50910e54e5", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found IQ1_S 29.644079264 GB, IQ1_M 32.732397728 GB, Q2_K split parts totaling 52.103164672 GB, IQ3_XS 58.228204288 GB, Q3_K_S 61.498188544 GB, Q3_K_M 67.789644544 GB, Q3_K_L 72.579801856 GB, IQ4_XS 76.354196224 GB, Q4_K_S 80.478098176 GB, Q4_K_M 85.586760448 GB, Q5_K_S 96.973705984 GB, Q5_K_M 99.968439040 GB, Q6_K 115.529728768 GB, Q8_0 149.414798080 GB, and FP16 split parts totaling 281.243411424 GB. The API gguf.totalFileSize exactly matches the IQ1_M file." }, { "label": "MaziyarPanahi Mixtral 8x22B v0.1 IQ1_M GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/resolve/053abe826eb9c5b1794e97bba22f0c50910e54e5/Mixtral-8x22B-v0.1.IQ1_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 25 metadata entries and 563 tensors. The linked file is 32.732397728 GB. Tensor spans sum to 32.731643904 GB; metadata/tokenizer/header/file overhead accounts for 0.000753824 GB. Tensor spans split into IQ1_M 28.824305664 GB, Q2_K 1.914200064 GB, Q5_K 1.588494336 GB, Q4_K 0.396361728 GB, F16 0.005505024 GB, and F32 0.002777088 GB. token_embd.weight is 0.064512 GB and resident-only; output.weight is 0.135168 GB and swept. Routed expert tensors sum to 30.211571712 GB, or 3.776446464 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 2.455560192 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, target config, pinned base config, linked GGUF file size checks, and a direct selected-IQ1_M GGUF header/tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a flat 2-bit GGUF and did not account for the API-selected IQ1_M artifact, mixed tensor classes, separate output projection, exact file overhead, or MoE routed expert byte groups." }, { "id": "maziyarpanahi--phi-3-5-mini-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Phi-3.5-mini-instruct-GGUF", "title": "MaziyarPanahi Phi-3.5 Mini Instruct GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the selected Q4_K_M GGUF artifact of Phi-3.5 Mini Instruct.", "model_family": "phi3-dense-gguf", "base_model_proof": { "base_model": "microsoft/Phi-3.5-mini-instruct", "relation": "quantized", "source": "Hugging Face API metadata, model card, selected GGUF header metadata, and audited Microsoft BF16 Phi-3.5 Mini profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of microsoft/Phi-3.5-mini-instruct. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q4_K_M GGUF header records the Phi3 architecture and matches the audited BF16 base geometry: 32 layers, 32 attention heads, 32 KV heads, 96 key/value head dimension, 3072 hidden size, 131072 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "phi-3-5-mini", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.821079648, "swept_params_b": 3.72257904, "auxiliary_resident_params_b": 0.098500608, "resident_weight_gb": 2.393232608, "swept_weight_gb": 2.33708736, "auxiliary_resident_weight_gb": 0.056145248, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Phi-3.5-mini-instruct.Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, rope factor tensors, and output.weight from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q4_K_M linked file is 2.393232608 GB. GGUF tensor payloads total 2.392493952 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000738656 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected GGUF adds two tiny RoPE factor tensors relative to the BF16 safetensors base profile." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 96, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 Phi3 decoder layers with 32 KV heads and 96-dimensional key/value heads. The header records phi3.attention.sliding_window 262144, above the 131072-token context length, so it does not cap this profile. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact. The repo contains many quantizations; those should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6263236646356344, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor payload bytes for swept weight traffic. Dequantization, llama.cpp kernels, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The selected artifact is Q4_K_M because the rendered model card usage examples select Q4_K_M for llama.cpp and other local runtimes. The HF API gguf.totalFileSize points at the smaller IQ1_M artifact, so that API field is recorded as mismatched and is not used as the selected artifact." }, "evidence": [ { "label": "MaziyarPanahi Phi-3.5 Mini Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "license", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "At commit 53fbed70b1fbf95c833c095523825cce68ebc063, the API records a public non-gated GGUF repo with base_model microsoft/Phi-3.5-mini-instruct, base_model:quantized, region:us, GGUF architecture phi3, context length 131072, gguf.total 3821079648, and current downloads 210087. The API gguf.totalFileSize is 917107424 bytes and points at Phi-3.5-mini-instruct.IQ1_M.gguf rather than the Q4_K_M artifact selected by the usage examples." }, { "label": "MaziyarPanahi Phi-3.5 Mini Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The card records model creator microsoft, original model microsoft/Phi-3.5-mini-instruct, quantized_by MaziyarPanahi, and standard GGUF runtime guidance. The rendered llama.cpp, Docker, Ollama, and Lemonade examples select Q4_K_M." }, { "label": "Microsoft Phi-3.5 Mini audited BF16 profile", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Phi3ForCausalLM geometry: 32 layers, 32 KV heads, 96 head dimension, 131072 max positions, full-context KV within that profile boundary, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Phi-3.5 Mini Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF/tree/53fbed70b1fbf95c833c095523825cce68ebc063", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "The repo tree and HEAD checks found IQ1_M 0.917107424 GB, Q4_K_M 2.393232608 GB, and Q8_0 4.061222624 GB. The selected Q4_K_M artifact intentionally differs from the API gguf.totalFileSize artifact because the rendered runtime examples choose Q4_K_M." }, { "label": "MaziyarPanahi Phi-3.5 Mini Instruct Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF/resolve/53fbed70b1fbf95c833c095523825cce68ebc063/Phi-3.5-mini-instruct.Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 40 metadata entries and 197 tensors. The selected file is 2.393232608 GB, with tensor payloads starting at byte 738656. Tensor payloads sum to 2.392493952 GB across 3.821079648B logical elements: token_embd.weight 0.055406592 GB, output.weight 0.080801280 GB, blk.* tensors 2.256273408 GB, output_norm.weight 0.000012288 GB, rope_factors_long.weight 0.000000192 GB, and rope_factors_short.weight 0.000000192 GB. Metadata/tokenizer/header/file overhead accounts for 0.000738656 GB. Stored tensor bytes split into Q4_K 1.357737984 GB, Q5_K 0.622854144 GB, Q6_K 0.411102720 GB, and F32 0.000799104 GB. The header records general.architecture phi3, MIT license, phi3.block_count 32, context_length 131072, embedding_length 3072, feed_forward_length 8192, attention.head_count 32, attention.head_count_kv 32, rope.dimension_count 96, and attention.sliding_window 262144." }, { "label": "MaziyarPanahi Phi-3.5 Mini Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF/raw/53fbed70b1fbf95c833c095523825cce68ebc063/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata, selected GGUF header, and audited BF16 base profile. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file size checks, stale repo-config check, and direct selected Q4_K_M GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit/mistral-style GGUF and undercounted both context/KV and selected Q4_K_M file bytes." }, { "id": "maziyarpanahi--phi-4-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/phi-4-GGUF", "title": "MaziyarPanahi Phi-4 GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Microsoft Phi-4.", "model_family": "phi4-dense-gguf", "base_model_proof": { "base_model": "microsoft/phi-4", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected FP16 GGUF header metadata, stale repo-config check, and audited Microsoft BF16 Phi-4 profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of microsoft/phi-4. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected FP16 GGUF header records the same Phi3ForCausalLM geometry as the audited BF16 base profile: 40 layers, hidden size 5120, intermediate size 17920, 40 attention heads, 10 KV heads, 16384 context, no sliding window, and untied embeddings with separate output.weight." }, "architecture": { "canonical_architecture_id": "phi-4", "max_context_tokens": 16384, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.6595072, "swept_params_b": 14.14570496, "auxiliary_resident_params_b": 0.51380224, "resident_weight_gb": 29.323399264, "swept_weight_gb": 28.29223936, "auxiliary_resident_weight_gb": 1.031159904, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for phi-4.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 29.323399264 GB. GGUF tensor spans total 29.319843840 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.003555424 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. GGUF stores small norm tensors as F32 while most matrices are F16, so exact tensor spans drive the bound." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 10, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 Phi-3/Phi-4 decoder layers, 10 KV heads, 128-dimensional key/value heads, 16384 context, and phi3.attention.sliding_window 0. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The card lists multiple lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, kernels, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is phi-4.fp16.gguf because gguf.totalFileSize exactly matches that linked object. It stores most tensors as F16 and small norm tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Phi-4 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/phi-4-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 7b3c692e11c2028dd77bd2c22ecc651fb1a32261, the API records a public non-gated GGUF text-generation repo with base_model microsoft/phi-4, base_model:quantized metadata, quantized tags, region:us, 112039 downloads, GGUF architecture phi3, context_length 16384, gguf.total 14659507200, and gguf.totalFileSize 29323399264." }, { "label": "MaziyarPanahi Phi-4 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/raw/7b3c692e11c2028dd77bd2c22ecc651fb1a32261/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model microsoft/phi-4, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base model record MIT licensing." }, { "label": "Microsoft Phi-4 audited BF16 base profile", "url": "https://huggingface.co/microsoft/phi-4", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The audited BF16 base profile records Phi3ForCausalLM, MIT licensing, 40 layers, hidden size 5120, intermediate size 17920, 40 attention heads, 10 KV heads, 128 key/value head dimension, 16384 max positions, no sliding window, and untied embeddings with separate model.embed_tokens.weight and lm_head.weight." }, { "label": "MaziyarPanahi Phi-4 GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/tree/7b3c692e11c2028dd77bd2c22ecc651fb1a32261", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF linked-object metadata found Q2_K 5.547348064 GB, Q3_K_S 6.504747104 GB, Q3_K_M 7.363268704 GB, Q3_K_L 7.930155104 GB, Q4_K_S 8.440762464 GB, Q4_K_M 9.053114464 GB, Q5_K_S 10.151579744 GB, Q5_K_M 10.604187744 GB, Q6_K 12.030251104 GB, Q8_0 15.580500064 GB, and fp16 29.323399264 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Phi-4 FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/resolve/7b3c692e11c2028dd77bd2c22ecc651fb1a32261/phi-4.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "license" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 33 metadata entries and 243 tensors. The linked file is 29.323399264 GB, with tensor data beginning at byte 3555424. Tensor spans sum to 29.319843840 GB: output.weight 1.027604480 GB, token_embd.weight 1.027604480 GB, blk.* tensors 27.264614400 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.003555424 GB. Stored tensor bytes split into F16 29.318184960 GB and F32 0.001658880 GB. The header records general.architecture phi3, MIT license, 40 blocks, 16384 context, 5120 embedding length, 17920 feed-forward length, 40 attention heads, 10 KV heads, key/value length 128, rope.freq_base 250000, sliding_window 0, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Phi-4 GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/raw/7b3c692e11c2028dd77bd2c22ecc651fb1a32261/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, which conflicts with the API and selected GGUF header. This profile records that check and intentionally uses the selected GGUF header plus audited BF16 base profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, expanded linked-file size metadata for all GGUF siblings, the stale repo config check, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Phi-4 FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or the stale repo-local Mistral config from this profile." }, { "id": "maziyarpanahi--phi-4-mini-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Phi-4-mini-instruct-GGUF", "title": "MaziyarPanahi Phi-4 Mini Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Phi-4 Mini Instruct.", "model_family": "phi4-mini-dense-gguf", "base_model_proof": { "base_model": "microsoft/Phi-4-mini-instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, and audited Microsoft BF16 Phi-4 Mini profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of microsoft/Phi-4-mini-instruct. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected FP16 GGUF header records the same Phi3ForCausalLM geometry as the audited BF16 base profile: 32 layers, hidden size 3072, intermediate size 8192, 24 attention heads, 8 KV heads, 131072 context, and tied input/output embeddings with no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "phi-4-mini", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.836021856, "swept_params_b": 3.836021856, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.680694592, "swept_weight_gb": 7.672443264, "auxiliary_resident_weight_gb": 0.008251328, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Phi-4-mini-instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors", "notes": "The selected FP16 linked file is 7.680694592 GB. Header tensor spans total 7.672443264 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.008251328 GB. The main GGUF contains token_embd.weight, blk.* tensors, output_norm.weight, and two tiny RoPE factor tensors, but no output.weight tensor; token_embd.weight is therefore charged as tied output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 Phi3 decoder layers, 8 KV heads, hidden size 3072, 24 attention heads, 131072 context, and phi3.attention.sliding_window 262144. The 128 key/value head dimension follows from hidden size divided by attention heads and matches the audited BF16 base profile. Because the sliding window is larger than the served context, this profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The repo contains many lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF. It stores most tensors as F16 and small norm/RoPE tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Phi-4 Mini Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Phi-4-mini-instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 177ca4b8b568a3f865c34e21e914aeb553da19cc, the API records a public non-gated GGUF text-generation repo with base_model microsoft/Phi-4-mini-instruct, base_model:quantized metadata, quantized tags, region:us, 126346 downloads, GGUF architecture phi3, context_length 131072, gguf.total 3836021856, and gguf.totalFileSize 7680694592." }, { "label": "MaziyarPanahi Phi-4 Mini Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model microsoft/Phi-4-mini-instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base model record MIT licensing." }, { "label": "Microsoft Phi-4 Mini audited BF16 profile", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records Phi3ForCausalLM, bfloat16, tied embeddings with no stored lm_head.weight, 32 layers, hidden size 3072, intermediate size 8192, 24 attention heads, 8 KV heads, 128 key/value head dimension, 131072 max positions, and full-context KV within that profile boundary." }, { "label": "MaziyarPanahi Phi-4 Mini Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF/tree/177ca4b8b568a3f865c34e21e914aeb553da19cc", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found Q2_K 1.682636096 GB, Q3_K_L 2.249653568 GB, Q3_K_M 2.117532992 GB, Q3_K_S 1.897332032 GB, Q4_K_M 2.491874624 GB, Q4_K_S 2.337733952 GB, Q5_K_M 2.848128320 GB, Q5_K_S 2.727804224 GB, Q6_K 3.155623232 GB, Q8_0 4.084611392 GB, and fp16 7.680694592 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Phi-4 Mini Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF/resolve/177ca4b8b568a3f865c34e21e914aeb553da19cc/Phi-4-mini-instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 36 metadata entries and 196 tensors. The linked file is 7.680694592 GB. Tensor spans sum to 7.672443264 GB: token_embd.weight 1.229193216 GB, blk.* tensors 6.443237376 GB, output_norm.weight 0.000012288 GB, rope_factors_long.weight 0.000000192 GB, and rope_factors_short.weight 0.000000192 GB. Metadata/tokenizer/header/file overhead accounts for 0.008251328 GB. Stored tensor bytes split into F16 7.671644160 GB and F32 0.000799104 GB. The header records general.architecture phi3, MIT license, phi3.block_count 32, context_length 131072, embedding_length 3072, feed_forward_length 8192, attention.head_count 24, attention.head_count_kv 8, rope.dimension_count 96, and attention.sliding_window 262144." }, { "label": "MaziyarPanahi Phi-4 Mini Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF/raw/177ca4b8b568a3f865c34e21e914aeb553da19cc/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, which conflicts with the API and selected GGUF header. This profile records that check and intentionally uses the selected GGUF header plus audited BF16 base profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, expanded linked-file size metadata for all GGUF siblings, the stale repo config check, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Phi-4 Mini Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or the stale repo-local Mistral config from this profile." }, { "id": "maziyarpanahi--qwen2-5-1-5b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF", "title": "MaziyarPanahi Qwen2.5 1.5B Instruct GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the catalog-selected Q2_K GGUF artifact of Qwen2.5 1.5B Instruct.", "model_family": "qwen2.5-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, stale repo-config check, pinned base config/API metadata, and linked-object size checks", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2.5-1.5B-Instruct. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q2_K GGUF header records the same dense Qwen2 tensor geometry as the pinned BF16 base config: 28 layers, hidden size 1536, intermediate size 8960, 12 attention heads, 2 KV heads, 128 key/value head dimension, 32768 context, and tied embeddings represented by the absence of a separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.543714304, "swept_params_b": 1.543714304, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.67630512, "swept_weight_gb": 0.67035392, "auxiliary_resident_weight_gb": 0.0059512, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen2.5-1.5B-Instruct.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because no output.weight tensor is stored and token_embd.weight is the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors", "notes": "The selected Q2_K linked file is 0.676305120 GB. GGUF tensor spans total 0.670353920 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005951200 GB. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen2 decoder layers, 2 KV heads, 128-dimensional key/value heads, and qwen2.context_length 32768. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the catalog-selected Q2_K GGUF artifact. FP16 and other quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.43810251563232255, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets Qwen2.5-1.5B-Instruct.Q2_K.gguf because the catalog row is the 2-bit GGUF entry. The live HF API gguf.totalFileSize currently matches the FP16 sibling, so exact selected-file header bytes are authoritative. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen2.5 1.5B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "At commit f67db4478e8de736c5b4f05eac67a9a84f1376a3, the API records a public non-gated GGUF text-generation repo with base_model Qwen/Qwen2.5-1.5B-Instruct, base_model:quantized metadata, quantized tags, region:us, 117624 downloads, GGUF architecture qwen2, context_length 32768, gguf.total 1543714304, and gguf.totalFileSize 3093669376. The API totalFileSize matches the fp16 sibling, while this profile targets the Q2_K file for the catalog's 2-bit row." }, { "label": "MaziyarPanahi Qwen2.5 1.5B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/raw/f67db4478e8de736c5b4f05eac67a9a84f1376a3/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The pinned card records original model Qwen/Qwen2.5-1.5B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It declares quantized 2/3/4/5/6/8-bit GGUF variants but does not include a detailed file-selection table." }, { "label": "Qwen2.5 1.5B Instruct base config and API metadata", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/989aa7980e4cf806f80c7fef2b1adb7bc71aa306/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The pinned base config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 1536, intermediate size 8960, 12 attention heads, 2 KV heads, 32768 max positions, 151936 vocabulary size, rope_theta 1000000, use_sliding_window false, and tied embeddings. The base API records Apache-2.0 licensing." }, { "label": "MaziyarPanahi Qwen2.5 1.5B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/tree/f67db4478e8de736c5b4f05eac67a9a84f1376a3", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found IQ1_S 0.436527840 GB, IQ1_M 0.464461536 GB, IQ2_XS 0.550327008 GB, Q2_K 0.676305120 GB, IQ3_XS 0.731699424 GB, Q3_K_S 0.760944864 GB, Q3_K_M 0.824178912 GB, Q3_K_L 0.880163040 GB, IQ4_XS 0.895731936 GB, Q4_K_S 0.940312800 GB, Q4_K_M 0.986048736 GB, Q5_K_S 1.098729696 GB, Q5_K_M 1.125050592 GB, Q6_K 1.272740064 GB, Q8_0 1.646573280 GB, and fp16 3.093669376 GB. The selected Q2_K artifact is the standard 2-bit GGUF file for this row." }, { "label": "MaziyarPanahi Qwen2.5 1.5B Instruct Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/resolve/f67db4478e8de736c5b4f05eac67a9a84f1376a3/Qwen2.5-1.5B-Instruct.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 16MB range-read of the selected GGUF v3 header found 38 metadata entries and 338 tensors. The linked file is 0.676305120 GB. Tensor spans sum to 0.670353920 GB across 1.543714304B logical elements: token_embd.weight 0.191439360 GB, blk.* tensors 0.478908416 GB, and output_norm.weight 0.000006144 GB. Metadata/tokenizer/header/file overhead accounts for 0.005951200 GB. Stored tensor spans split into Q2_K 0.278175744 GB, Q3_K 0.193966080 GB, Q4_K 0.006193152 GB, Q6_K 0.191439360 GB, and F32 0.000579584 GB. The header records general.architecture qwen2, Apache-2.0 license, qwen2.block_count 28, context_length 32768, embedding_length 1536, feed_forward_length 8960, attention.head_count 12, attention.head_count_kv 2, rope.freq_base 1000000, and no separate output.weight tensor." }, { "label": "MaziyarPanahi Qwen2.5 1.5B Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/raw/f67db4478e8de736c5b4f05eac67a9a84f1376a3/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base config and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned base config/API metadata, linked GGUF file sizes, stale repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the row-selected Qwen2.5 1.5B Instruct Q2_K GGUF artifact from MaziyarPanahi. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--qwen2-5-7b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF", "title": "MaziyarPanahi Qwen2.5 7B Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Qwen2.5 7B Instruct.", "model_family": "qwen2.5-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, and audited BF16 Qwen2.5 7B Instruct base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2.5-7B-Instruct. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected FP16 GGUF header records the same Qwen2 tensor geometry as the audited BF16 base profile: 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128 key/value head dimension, separate input/output embeddings, and 32768 context." }, "architecture": { "canonical_architecture_id": "qwen2-5-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.237853632, "swept_weight_gb": 14.141904896, "auxiliary_resident_weight_gb": 1.095948736, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen2.5-7B-Instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 15.237853632 GB. GGUF tensor spans total 15.231899648 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005953984 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen2 decoder layers, 4 KV heads, 128-dimensional key/value heads, and qwen2.context_length 32768. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The card lists multiple lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF. It stores most tensors as F16 and small norm/bias tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen2.5 7B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 113085215169f69e309fa57d351d7e2b06e1350e, the API records a public non-gated GGUF text-generation repo with base_model Qwen/Qwen2.5-7B-Instruct, base_model:quantized metadata, quantized tags, region:us, 124355 downloads, GGUF architecture qwen2, context_length 32768, gguf.total 7615616512, and gguf.totalFileSize 15237853632." }, { "label": "MaziyarPanahi Qwen2.5 7B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen2.5-7B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen2.5 7B Instruct audited BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128 head dimension, 32768 max positions, separate stored model.embed_tokens.weight plus lm_head.weight, and use_sliding_window false." }, { "label": "MaziyarPanahi Qwen2.5 7B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF/tree/113085215169f69e309fa57d351d7e2b06e1350e", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found fp16 15.237853632 GB, Q8_0 8.098525856 GB, Q6_K 6.254199456 GB, Q5_K_M 5.444831904 GB, Q5_K_S 5.315177120 GB, Q4_K_M 4.683074208 GB, Q4_K_S 4.457769632 GB, IQ4_XS 4.218473120 GB, Q3_K_L 4.088459936 GB, Q3_K_M 3.808391840 GB, Q3_K_S 3.492369056 GB, IQ3_XS 3.346256544 GB, Q2_K 3.015940768 GB, IQ2_XS 2.469022368 GB, IQ1_M 2.042196640 GB, and IQ1_S 1.903667872 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen2.5 7B Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF/resolve/113085215169f69e309fa57d351d7e2b06e1350e/Qwen2.5-7B-Instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 34 metadata entries and 339 tensors. The linked file is 15.237853632 GB. Tensor spans sum to 15.231899648 GB: output.weight 1.089994752 GB, token_embd.weight 1.089994752 GB, blk.* tensors 13.051895808 GB, and output_norm.weight 0.000014336 GB. Metadata/tokenizer/header/file overhead accounts for 0.005953984 GB. Stored tensor bytes split into F16 15.230566400 GB and F32 0.001333248 GB. The header records general.architecture qwen2, qwen2.block_count 28, context_length 32768, embedding_length 3584, feed_forward_length 18944, attention.head_count 28, attention.head_count_kv 4, rope.freq_base 1000000, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen2.5 7B Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF/raw/113085215169f69e309fa57d351d7e2b06e1350e/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, which conflicts with the API and selected GGUF header. This profile records that check and intentionally uses the selected GGUF header plus audited base profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, expanded linked-file size metadata for all GGUF siblings, the stale repo config check, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Qwen2.5 7B Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or the stale repo-local Mistral config from this profile." }, { "id": "maziyarpanahi--qwen2-7b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen2-7B-Instruct-GGUF", "title": "MaziyarPanahi Qwen2 7B Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Qwen2 7B Instruct.", "model_family": "qwen2-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2-7B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, base model API/config metadata, and audited BF16 Qwen2 7B Instruct base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2-7B-Instruct. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected FP16 GGUF header records the same Qwen2 tensor geometry as the audited BF16 base profile: 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128 key/value head dimension, separate input/output embeddings, and 32768 context." }, "architecture": { "canonical_architecture_id": "qwen2-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.237850656, "swept_weight_gb": 14.141904896, "auxiliary_resident_weight_gb": 1.09594576, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen2-7B-Instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 15.237850656 GB. GGUF tensor spans total 15.231899648 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005951008 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen2 decoder layers, 4 KV heads, 128-dimensional key/value heads, and qwen2.context_length 32768. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The card lists multiple lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF. It stores most tensors as F16 and small norm/bias tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen2 7B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2-7B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 8cd289397d4eb8a1faa022adb9bbab5aae2d1739, the API records a public non-gated GGUF text-generation repo with base_model Qwen/Qwen2-7B-Instruct, base_model:quantized metadata, quantized tags, region:us, 113445 downloads, GGUF architecture qwen2, context_length 32768, gguf.total 7615616512, and gguf.totalFileSize 15237850656." }, { "label": "MaziyarPanahi Qwen2 7B Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/raw/8cd289397d4eb8a1faa022adb9bbab5aae2d1739/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen2-7B-Instruct, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown even though the base model is Apache-2.0." }, { "label": "Qwen2 7B Instruct base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-7B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit f2826a00ceef68f0f2b946d945ecc0477ce4450c, the base repo is public, Apache-2.0 licensed, text-generation tagged, and records BF16 safetensors total 7615616512 parameters." }, { "label": "Qwen2 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/raw/f2826a00ceef68f0f2b946d945ecc0477ce4450c/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The base config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 32768 max positions, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, vocab size 152064, and RoPE theta 1000000." }, { "label": "MaziyarPanahi Qwen2 7B Instruct GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/tree/8cd289397d4eb8a1faa022adb9bbab5aae2d1739", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found fp16 15.237850656 GB, Q8_0 8.098522880 GB, Q6_K 6.254196480 GB, Q5_K_M 5.444828928 GB, Q5_K_S 5.315174144 GB, Q4_K_M 4.683071232 GB, Q4_K_S 4.457766656 GB, IQ4_XS 4.218470144 GB, Q3_K_L 4.088456960 GB, Q3_K_M 3.808388864 GB, Q3_K_S 3.492366080 GB, IQ3_XS 3.346253568 GB, Q2_K 3.015937792 GB, IQ2_XS 2.469019392 GB, IQ1_M 2.042193664 GB, and IQ1_S 1.903664896 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen2 7B Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/resolve/8cd289397d4eb8a1faa022adb9bbab5aae2d1739/Qwen2-7B-Instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 21 metadata entries and 339 tensors. The linked file is 15.237850656 GB, with tensor data beginning at byte 5951008. Tensor spans sum to 15.231899648 GB: output.weight 1.089994752 GB, token_embd.weight 1.089994752 GB, blk.* tensors 13.051895808 GB, and output_norm.weight 0.000014336 GB. Metadata/tokenizer/header/file overhead accounts for 0.005951008 GB. Stored tensor bytes split into F16 15.230566400 GB and F32 0.001333248 GB. The header records general.architecture qwen2, qwen2.block_count 28, context_length 32768, embedding_length 3584, feed_forward_length 18944, attention.head_count 28, attention.head_count_kv 4, rope.freq_base 1000000, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen2 7B Instruct GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/raw/8cd289397d4eb8a1faa022adb9bbab5aae2d1739/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, which conflicts with the API and selected GGUF header. This profile records that check and intentionally uses the selected GGUF header plus base config as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, expanded linked-file size metadata for all GGUF siblings, the stale repo config check, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Qwen2 7B Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or the stale repo-local Mistral config from this profile." }, { "id": "maziyarpanahi--qwen3-0-6b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-0.6B-GGUF", "title": "MaziyarPanahi Qwen3 0.6B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 0.6B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 0.6B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-0.6B. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 28 layers, 16 attention heads, 8 KV heads, 128 key/value head dimension, 1024 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-0-6b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.751632384, "swept_params_b": 0.59604992, "auxiliary_resident_params_b": 0.155582464, "resident_weight_gb": 1.509347456, "swept_weight_gb": 1.192230912, "auxiliary_resident_weight_gb": 0.317116544, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-0.6B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 1.509347456 GB. GGUF tensor spans total 1.503395840 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005951616 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 0.6B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-0.6B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 16d75108d73a476af91a4f6df4cd77e854b42d04, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-0.6B, base_model:quantized, quantized tags, region:us, 231812 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 751632384, and gguf.totalFileSize 1509347456." }, { "label": "MaziyarPanahi Qwen3 0.6B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-0.6B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 0.6B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 28 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 0.6B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF/tree/16d75108d73a476af91a4f6df4cd77e854b42d04", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 1.509347456 GB, Q6_K 0.622733440 GB, Q5_K_M 0.551378048 GB, Q4_K_M 0.484220032 GB, Q3_K_L 0.435343488 GB, Q3_K_M 0.413978752 GB, and Q2_K 0.347288704 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 0.6B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF/resolve/16d75108d73a476af91a4f6df4cd77e854b42d04/Qwen3-0.6B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "An 8MB range-read of the selected GGUF v3 header found 34 metadata entries and 311 tensors. The linked file is 1.509347456 GB. Tensor spans sum to 1.503395840 GB: output.weight 0.311164928 GB, token_embd.weight 0.311164928 GB, blk.* tensors 0.881061888 GB, and output_norm.weight 0.000004096 GB. Metadata/tokenizer/header/file overhead accounts for 0.005951616 GB. Stored tensor bytes split into F16 1.503133696 GB and F32 0.000262144 GB. The header records general.architecture qwen3, qwen3.block_count 28, context_length 40960, embedding_length 1024, feed_forward_length 3072, attention.head_count 16, attention.head_count_kv 8, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen3 0.6B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF/raw/16d75108d73a476af91a4f6df4cd77e854b42d04/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, which conflicts with the API and selected GGUF header. This profile records that check and intentionally uses the selected GGUF header plus audited base profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked-object HEAD checks for all GGUF siblings, the stale repo config check, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Qwen3 0.6B F16 GGUF artifact. Do not infer architecture from the stale repo-local config.json or from the lower-bit GGUF siblings." }, { "id": "maziyarpanahi--qwen3-1-7b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-1.7B-GGUF", "title": "MaziyarPanahi Qwen3 1.7B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 1.7B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-1.7B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 1.7B profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-1.7B. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 parent geometry: 28 layers, 16 attention heads, 8 KV heads, 128 key/value head dimension, 2048 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-1-7b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.031739904, "swept_params_b": 1.720574976, "auxiliary_resident_params_b": 0.311164928, "resident_weight_gb": 4.069679232, "swept_weight_gb": 3.44139776, "auxiliary_resident_weight_gb": 0.628281472, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-1.7B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 4.069679232 GB. GGUF tensor spans total 4.063727616 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005951616 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. GGUF stores output_norm and layer norms as F32, so tensor bytes are slightly larger than the BF16 safetensors parent." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0030512882026854, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 1.7B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-1.7B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 57a69d3bba4c7254d3466a3bb96d909b988f8fbd, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-1.7B, base_model:quantized, quantized tags, region:us, 224978 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 2031739904, and gguf.totalFileSize 4069679232." }, { "label": "MaziyarPanahi Qwen3 1.7B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/raw/57a69d3bba4c7254d3466a3bb96d909b988f8fbd/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-1.7B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 1.7B audited parent profile", "url": "https://huggingface.co/Qwen/Qwen3-1.7B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 parent profile records the same Qwen3ForCausalLM geometry: 28 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 1.7B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/tree/57a69d3bba4c7254d3466a3bb96d909b988f8fbd", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 4.069679232 GB, Q6_K 1.673007232 GB, Q5_K_M 1.471805568 GB, Q4_K_M 1.282439296 GB, Q3_K_L 1.137205376 GB, Q3_K_M 1.073242240 GB, and Q2_K 0.879896704 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 1.7B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/resolve/57a69d3bba4c7254d3466a3bb96d909b988f8fbd/Qwen3-1.7B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 34 metadata entries and 311 tensors. The linked file is 4.069679232 GB. Tensor spans sum to 4.063727616 GB: output.weight 0.622329856 GB, token_embd.weight 0.622329856 GB, blk.* tensors 2.819059712 GB, and output_norm.weight 0.000008192 GB. Metadata/tokenizer/header/file overhead accounts for 0.005951616 GB. Stored tensor bytes split into F16 4.063232000 GB and F32 0.000495616 GB. The header records general.architecture qwen3, qwen3.block_count 28, context_length 40960, embedding_length 2048, feed_forward_length 6144, attention.head_count 16, attention.head_count_kv 8, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen3 1.7B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/raw/57a69d3bba4c7254d3466a3bb96d909b988f8fbd/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 parent profile, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit GGUF and undercounted the API-selected F16 artifact." }, { "id": "maziyarpanahi--qwen3-14b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-14B-GGUF", "title": "MaziyarPanahi Qwen3 14B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 14B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-14B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 14B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-14B. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 40 layers, 40 attention heads, 8 KV heads, 128 key/value head dimension, 5120 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-14b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7683072, "swept_params_b": 13.99039488, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 29.543423744, "swept_weight_gb": 27.98163968, "auxiliary_resident_weight_gb": 1.561784064, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-14B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 29.543423744 GB. GGUF tensor spans total 29.537464320 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005959424 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. GGUF stores output_norm and layer norms as F32, so tensor bytes are slightly larger than the BF16 safetensors base." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.000461078166088, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 14B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-14B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 421b77e6e553db95b5ca5eae8055b2677bc4f41f, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-14B, base_model:quantized, quantized tags, region:us, 229212 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 14768307200, and gguf.totalFileSize 29543423744." }, { "label": "MaziyarPanahi Qwen3 14B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-14B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 14B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-14B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 40 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 14B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF/tree/421b77e6e553db95b5ca5eae8055b2677bc4f41f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 29.543423744 GB, Q6_K 12.121937664 GB, Q5_K_M 10.514569984 GB, Q4_K_M 9.001753344 GB, Q3_K_L 7.900651264 GB, Q3_K_M 7.321313024 GB, and Q2_K 5.753983744 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 14B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF/resolve/421b77e6e553db95b5ca5eae8055b2677bc4f41f/Qwen3-14B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 34 metadata entries and 443 tensors. The linked file is 29.543423744 GB. Tensor spans sum to 29.537464320 GB: output.weight 1.555824640 GB, token_embd.weight 1.555824640 GB, blk.* tensors 26.425794560 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.005959424 GB. Stored tensor bytes split into F16 29.535764480 GB and F32 0.001699840 GB. The header records general.architecture qwen3, qwen3.block_count 40, context_length 40960, embedding_length 5120, feed_forward_length 17408, attention.head_count 40, attention.head_count_kv 8, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen3 14B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF/raw/421b77e6e553db95b5ca5eae8055b2677bc4f41f/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit GGUF and undercounted the API-selected F16 artifact." }, { "id": "maziyarpanahi--qwen3-30b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-30B-A3B-GGUF", "title": "MaziyarPanahi Qwen3 30B A3B GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Qwen3 30B A3B.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 30B A3B profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-30B-A3B. The repo has no config.json, so this profile uses the selected GGUF header and the audited BF16 Qwen3 30B A3B profile as architecture evidence. The selected GGUF header records the Qwen3Moe architecture and matches the audited BF16 parent geometry: 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 routed experts per token, no shared expert, and 40960 context tokens." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b", "max_context_tokens": 40960, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 11.2586096, "main_resident_weight_gb": 11.150538752, "auxiliary_resident_weight_gb": 0.108070848, "fixed_weight_gb": 0.656390144, "routed_expert_weight_gb": 0.081985536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3-30B-A3B.Q2_K.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.47 non-routed tensors, routers, and expected-distinct routed expert tensor groups from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "Header-derived stored bytes are used instead of rounded 30B/3B model-card parameters. The selected main GGUF mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 128 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata and audited BF16 config record 48 full-context attention layers with 4 KV heads and 128 key/value dimensions. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q2_K GGUF artifact. The repo contains several larger quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3687463770091794, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and any runtime-specific expert routing locality are outside Bounds Engine v1.", "notes": "The API-selected artifact is Qwen3-30B-A3B.Q2_K.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 30B A3B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-30B-A3B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 6443554b0ef1176268cae69212105bf4c7187091, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-30B-A3B, base_model:quantized, quantized tags, region:us, 222592 downloads, GGUF architecture qwen3moe, 40960 context length, gguf.total 30532122624, and gguf.totalFileSize 11258609600." }, { "label": "MaziyarPanahi Qwen3 30B A3B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/raw/6443554b0ef1176268cae69212105bf4c7187091/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-30B-A3B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 30B A3B audited parent profile", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The existing audited BF16 profile records Qwen3MoeForCausalLM, qwen3_moe, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 40960, and vocab_size 151936." }, { "label": "MaziyarPanahi Qwen3 30B A3B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/tree/6443554b0ef1176268cae69212105bf4c7187091", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q6_K 25.092532160 GB, Q5_K_M 21.725581248 GB, Q4_K_M 18.556686272 GB, Q3_K_L 15.900669888 GB, Q3_K_M 14.711846848 GB, and Q2_K 11.258609600 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 30B A3B Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/resolve/6443554b0ef1176268cae69212105bf4c7187091/Qwen3-30B-A3B.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 37 metadata entries and 579 tensors. The linked file is 11.258609600 GB. Tensor spans sum to 11.252639744 GB; metadata/tokenizer/header/file overhead accounts for 0.005969856 GB. Tensor spans split into Q2_K 6.592524288 GB, Q3_K 4.325376000 GB, Q6_K 0.255252480 GB, F32 0.051175424 GB, and Q4_K 0.028311552 GB. token_embd.weight is 0.102100992 GB and resident-only; output.weight is 0.255252480 GB and swept. Routed expert tensors sum to 10.494148608 GB, or 0.081985536 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 0.656390144 GB." }, { "label": "MaziyarPanahi Qwen3 30B A3B GGUF missing config check", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/raw/6443554b0ef1176268cae69212105bf4c7187091/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json returns 404, and the API config object is empty. This profile intentionally uses the selected GGUF header plus audited BF16 parent profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 parent profile, linked GGUF file sizes, missing repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a flat 2-bit GGUF and did not account for mixed tensor classes, separate output projection, exact file overhead, or MoE routed expert byte groups." }, { "id": "maziyarpanahi--qwen3-32b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-32B-GGUF", "title": "MaziyarPanahi Qwen3 32B GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Qwen3 32B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-32B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 32B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-32B. The repo has no config.json, so this profile uses the selected GGUF header and the audited BF16 Qwen3 32B profile as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 64 layers, 64 attention heads, 8 KV heads, 128 key/value head dimension, 5120 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-32b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.762123264, "swept_params_b": 31.984210944, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 12.344651584, "swept_weight_gb": 12.083424256, "auxiliary_resident_weight_gb": 0.261227328, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-32B.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 12.344651584 GB. GGUF tensor spans total 12.338676736 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005974848 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 64 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. The repo contains several larger quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.37679644522809885, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the Q2_K GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 32B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-32B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 0dd9f2cba883d7114816444ae9ee9219cf87a917, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-32B, base_model:quantized, quantized tags, region:us, 225531 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 32762123264, and gguf.totalFileSize 12344651584." }, { "label": "MaziyarPanahi Qwen3 32B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF/raw/0dd9f2cba883d7114816444ae9ee9219cf87a917/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-32B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 32B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-32B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 64 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 32B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF/tree/0dd9f2cba883d7114816444ae9ee9219cf87a917", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q6_K 26.883306304 GB, Q5_K_M 23.214831424 GB, Q4_K_M 19.762149184 GB, Q3_K_L 17.330993984 GB, Q3_K_M 15.971777344 GB, and Q2_K 12.344651584 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize. No F16 sibling exists in this repo." }, { "label": "MaziyarPanahi Qwen3 32B Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF/resolve/0dd9f2cba883d7114816444ae9ee9219cf87a917/Qwen3-32B.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 30 metadata entries and 707 tensors. The linked file is 12.344651584 GB. Tensor spans sum to 12.338676736 GB: output.weight 0.638131200 GB, token_embd.weight 0.255252480 GB, blk.* tensors 11.445272576 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.005974848 GB. Stored tensor bytes split into Q2_K 6.751180800 GB, Q3_K 4.757913600 GB, Q4_K 0.188743680 GB, Q6_K 0.638131200 GB, and F32 0.002707456 GB. The header records general.architecture qwen3, qwen3.block_count 64, context_length 40960, embedding_length 5120, feed_forward_length 25600, attention.head_count 64, attention.head_count_kv 8, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen3 32B GGUF missing config check", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF/raw/0dd9f2cba883d7114816444ae9ee9219cf87a917/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json returns 404, and the API config object is empty. This profile intentionally uses the selected GGUF header plus audited BF16 base profile as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, missing repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a flat 2-bit GGUF and did not account for mixed tensor classes, separate output projection, or exact file overhead." }, { "id": "maziyarpanahi--qwen3-4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-4B-GGUF", "title": "MaziyarPanahi Qwen3 4B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 4B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-4B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 4B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-4B. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 36 layers, 32 attention heads, 8 KV heads, 128 key/value head dimension, 2560 hidden size, 40960 context, and tied embeddings with no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "qwen3-4b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.022468096, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0, "resident_weight_gb": 8.051285152, "swept_weight_gb": 8.045328384, "auxiliary_resident_weight_gb": 0.005956768, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-4B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans because token_embd.weight is the tied output projection and no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors", "notes": "The selected F16 linked file is 8.051285152 GB. GGUF tensor spans total 8.045328384 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005956768 GB. Because no output.weight tensor is stored, token_embd.weight is charged as ordinary tied output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 36 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0015783742340463, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 4B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-4B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit b507d8f2622a6be905066595a5b3cd00e11ab379, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-4B, base_model:quantized, quantized tags, region:us, 226707 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 4022468096, and gguf.totalFileSize 8051285152." }, { "label": "MaziyarPanahi Qwen3 4B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/raw/b507d8f2622a6be905066595a5b3cd00e11ab379/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-4B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 4B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 36 layers, 8 KV heads, 128 head dimension, 40960 max positions, tied embeddings, and no separate lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/tree/b507d8f2622a6be905066595a5b3cd00e11ab379", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 8.051285152 GB, Q6_K 3.306261152 GB, Q5_K_M 2.889513632 GB, Q4_K_M 2.497280672 GB, Q3_K_L 2.239785632 GB, Q3_K_M 2.075617952 GB, and Q2_K 1.669499552 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 4B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/resolve/b507d8f2622a6be905066595a5b3cd00e11ab379/Qwen3-4B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 34 metadata entries and 398 tensors. The linked file is 8.051285152 GB. Tensor spans sum to 8.045328384 GB: token_embd.weight 0.777912320 GB, blk.* tensors 7.267405824 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.005956768 GB. Stored tensor bytes split into F16 8.044544000 GB and F32 0.000784384 GB. The header records general.architecture qwen3, qwen3.block_count 36, context_length 40960, embedding_length 2560, feed_forward_length 9728, attention.head_count 32, attention.head_count_kv 8, key/value head length 128, and no output.weight tensor." }, { "label": "MaziyarPanahi Qwen3 4B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/raw/b507d8f2622a6be905066595a5b3cd00e11ab379/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit GGUF and undercounted the API-selected F16 artifact." }, { "id": "maziyarpanahi--qwen3-4b-instruct-2507-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF", "title": "MaziyarPanahi Qwen3 4B Instruct 2507 GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 4B Instruct 2507.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-4B-Instruct-2507", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 4B Instruct 2507 base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-4B-Instruct-2507. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 36 layers, 32 attention heads, 8 KV heads, 128 key/value head dimension, 2560 hidden size, 262144 context, and tied token embeddings with no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "qwen3-4b-instruct-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.022468096, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0, "resident_weight_gb": 8.051284928, "swept_weight_gb": 8.045328384, "auxiliary_resident_weight_gb": 0.005956544, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for Qwen3-4B-Instruct-2507.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The profile targets Qwen3-4B-Instruct-2507.fp16.gguf because the live HF API gguf.totalFileSize matches that linked object. Header tensor spans total 8.045328384 GB, while the linked file size is 8.051284928 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight. It has no output.weight, mmproj, vision, audio, MTP, or draft tensor." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records full-context Qwen3 attention geometry with 36 layers, 8 KV heads, and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The repo contains multiple quantized GGUF siblings; those should get separate selected-artifact profiles if used." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0015783185468425, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 4B Instruct 2507 GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit aec29f0e8c31130ba811bec2c774c2ef44888f55, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-4B-Instruct-2507, base_model:quantized, quantized tags, region:us, 149675 downloads, GGUF architecture qwen3, 262144 context length, gguf.total 4022468096, and gguf.totalFileSize 8051284928." }, { "label": "MaziyarPanahi Qwen3 4B Instruct 2507 GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-4B-Instruct-2507, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 4B Instruct 2507 audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 36 layers, 8 KV heads, 128 head dimension, 262144 max positions, tied embeddings, and no separate lm_head.weight tensor." }, { "label": "MaziyarPanahi Qwen3 4B Instruct 2507 GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF/tree/aec29f0e8c31130ba811bec2c774c2ef44888f55", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 8.051284928 GB, Q6_K 3.306260928 GB, Q5_K_M 2.889513408 GB, Q4_K_M 2.497280448 GB, Q3_K_L 2.239785408 GB, Q3_K_M 2.075617728 GB, and Q2_K 1.669499328 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 4B Instruct 2507 F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF/resolve/aec29f0e8c31130ba811bec2c774c2ef44888f55/Qwen3-4B-Instruct-2507.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 32 metadata entries and 398 tensors. The linked file is 8.051284928 GB. Tensor spans sum to 8.045328384 GB: token_embd.weight 0.777912320 GB, blk.* tensors 7.267405824 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.005956544 GB. Stored tensor bytes split into F16 8.044544000 GB and F32 0.000784384 GB. The header records general.architecture qwen3, qwen3.block_count 36, context_length 262144, embedding_length 2560, feed_forward_length 9728, attention.head_count 32, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 5000000, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." }, { "label": "MaziyarPanahi Qwen3 4B Instruct 2507 GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF/raw/aec29f0e8c31130ba811bec2c774c2ef44888f55/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit GGUF and undercounted the API-selected F16 artifact." }, { "id": "maziyarpanahi--qwen3-8b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Qwen3-8B-GGUF", "title": "MaziyarPanahi Qwen3 8B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Qwen3 8B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-8B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen3 8B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-8B. The repo-level config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected GGUF header records the Qwen3 architecture and matches the audited BF16 base geometry: 36 layers, 32 attention heads, 8 KV heads, 128 key/value head dimension, 4096 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 16.388044, "swept_weight_gb": 15.137427456, "auxiliary_resident_weight_gb": 1.250616544, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-8B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 16.388044000 GB. GGUF tensor spans total 16.382087168 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005956832 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. GGUF stores output_norm and layer norms as F32, so tensor bytes are slightly larger than the BF16 safetensors base." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 36 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.000802526233737, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Qwen3 8B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-8B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit ac6dd95cf227fe9138362c0536fe3c3802008ccf, the API records a public non-gated GGUF repo with base_model Qwen/Qwen3-8B, base_model:quantized, quantized tags, region:us, 227628 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 8190735360, and gguf.totalFileSize 16388044000." }, { "label": "MaziyarPanahi Qwen3 8B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/raw/ac6dd95cf227fe9138362c0536fe3c3802008ccf/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/Qwen3-8B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, so the catalog license remains unknown." }, { "label": "Qwen3 8B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-8B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 36 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "MaziyarPanahi Qwen3 8B GGUF linked-object HEAD checks", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/tree/ac6dd95cf227fe9138362c0536fe3c3802008ccf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 16.388044000 GB, Q6_K 6.725899488 GB, Q5_K_M 5.851112672 GB, Q4_K_M 5.027783904 GB, Q3_K_L 4.431394016 GB, Q3_K_M 4.124161248 GB, and Q2_K 3.281732832 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Qwen3 8B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/resolve/ac6dd95cf227fe9138362c0536fe3c3802008ccf/Qwen3-8B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 34 metadata entries and 399 tensors. The linked file is 16.388044000 GB. Tensor spans sum to 16.382087168 GB: output.weight 1.244659712 GB, token_embd.weight 1.244659712 GB, blk.* tensors 13.892751360 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.005956832 GB. Stored tensor bytes split into F16 16.380854272 GB and F32 0.001232896 GB. The header records general.architecture qwen3, qwen3.block_count 36, context_length 40960, embedding_length 4096, feed_forward_length 12288, attention.head_count 32, attention.head_count_kv 8, and key/value head length 128." }, { "label": "MaziyarPanahi Qwen3 8B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/raw/ac6dd95cf227fe9138362c0536fe3c3802008ccf/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the API GGUF metadata and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a 2-bit GGUF and undercounted the API-selected F16 artifact." }, { "id": "maziyarpanahi--qwq-32b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/QwQ-32B-GGUF", "title": "MaziyarPanahi QwQ 32B GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Qwen QwQ 32B.", "model_family": "qwen2-dense-gguf", "base_model_proof": { "base_model": "Qwen/QwQ-32B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, package-config absence check, and pinned Qwen base config", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/QwQ-32B. The pinned Qwen base config records Qwen2ForCausalLM with 64 layers, hidden size 5120, intermediate size 27648, 40 attention heads, 8 KV heads, untied embeddings, and 40960 max positions. The selected GGUF header preserves the main tensor geometry but records general.architecture llama and llama.context_length 131072. This profile therefore uses the selected GGUF header as the served artifact source and does not claim full config compatibility." }, "architecture": { "canonical_architecture_id": "qwen-qwq-32b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 12.313707552, "swept_weight_gb": 12.051652608, "auxiliary_resident_weight_gb": 0.262054944, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for QwQ-32B.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 12.313707552 GB. GGUF tensor spans total 12.307120128 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.006587424 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 64 decoder blocks, 8 KV heads, 128-dimensional key/value heads, and llama.context_length 131072. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. The pinned base config records use_sliding_window false; the selected GGUF header does not expose a sliding-window field, so no sliding-window KV cap is applied." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.37583182831320683, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, kernels, scheduler behavior, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets QwQ-32B.Q2_K.gguf because the live HF API gguf.totalFileSize exactly matches that linked object. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi QwQ 32B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/QwQ-32B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit c289541240c414b14fc9a8b2d48fac949dc16012, the API records a public non-gated GGUF text-generation repo with base_model Qwen/QwQ-32B, base_model:quantized metadata, quantized tags, region:us, 112246 downloads, GGUF architecture llama, context_length 131072, gguf.total 32763876352, and gguf.totalFileSize 12313707552. The API totalFileSize matches QwQ-32B.Q2_K.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi QwQ 32B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/raw/c289541240c414b14fc9a8b2d48fac949dc16012/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model Qwen/QwQ-32B, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF header and base model record Apache-2.0 licensing." }, { "label": "Qwen QwQ 32B base API metadata", "url": "https://huggingface.co/api/models/Qwen/QwQ-32B", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit 976055f8c83f394f35dbd3ab09a285a984907bd0, the base repo is public, Apache-2.0 licensed, text-generation tagged, region:us tagged, and records BF16 safetensors total 32763876352 parameters." }, { "label": "Qwen QwQ 32B base config", "url": "https://huggingface.co/Qwen/QwQ-32B/raw/976055f8c83f394f35dbd3ab09a285a984907bd0/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The base config records Qwen2ForCausalLM, bfloat16, 64 layers, hidden size 5120, intermediate size 27648, 40 attention heads, 8 KV heads, 40960 max positions, sliding_window 32768, use_sliding_window false, tie_word_embeddings false, vocab size 152064, and RoPE theta 1000000." }, { "label": "MaziyarPanahi QwQ 32B GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/tree/c289541240c414b14fc9a8b2d48fac949dc16012", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF linked-object metadata found Q2_K 12.313707552 GB, Q3_K_S 14.392939552 GB, Q3_K_M 15.935656992 GB, Q3_K_L 17.247687712 GB, Q4_K_S 18.785018912 GB, Q4_K_M 19.851944992 GB, Q5_K_S 22.638863392 GB, Q5_K_M 23.262766112 GB, Q6_K 26.886763552 GB, and Q8_0 34.821493792 GB. The selected Q2_K artifact exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi QwQ 32B Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/resolve/c289541240c414b14fc9a8b2d48fac949dc16012/QwQ-32B.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "license" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 34 metadata entries and 771 tensors. The linked file is 12.313707552 GB, with tensor data beginning at byte 6587424. Tensor spans sum to 12.307120128 GB: output.weight 0.638668800 GB, output_norm.weight 0.000020480 GB, token_embd.weight 0.255467520 GB, and blk.* tensors 11.412963328 GB. Metadata/tokenizer/header/file overhead accounts for 0.006587424 GB. Stored tensor bytes split into Q2_K 6.861496320 GB, Q3_K 4.613734400 GB, Q4_K 0.188743680 GB, Q6_K 0.638668800 GB, and F32 0.004476928 GB. The header records general.architecture llama, Apache-2.0 license, 64 blocks, 131072 context, 5120 embedding length, 27648 feed-forward length, 40 attention heads, 8 KV heads, 128 RoPE/head dimension, RoPE freq base 1000000, 152064 vocab size, and separate output.weight." }, { "label": "MaziyarPanahi QwQ 32B GGUF package config absence", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/raw/c289541240c414b14fc9a8b2d48fac949dc16012/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The package has no repo-local config.json; the raw config request returned HTTP 404. This profile therefore uses the selected GGUF header plus the pinned Qwen base config as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, expanded linked-file size metadata for all GGUF siblings, package config absence check, and a direct GGUF header/tensor-index range read of the selected Q2_K artifact." }, "notes": "Use this profile for the API-selected QwQ 32B Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--solar-pro-preview-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/solar-pro-preview-instruct-GGUF", "title": "MaziyarPanahi Solar Pro Preview Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Solar Pro Preview Instruct.", "model_family": "solar-pro-dense-gguf", "base_model_proof": { "base_model": "upstage/solar-pro-preview-instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, selected linked-object size checks, base config, and Solar custom-code review", "config_compatible": false, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of upstage/solar-pro-preview-instruct. The selected GGUF header matches the base layer, hidden, feed-forward, head, KV-head, vocabulary, context, and untied embedding geometry, but the base config records sliding_window 2047 while the selected GGUF header does not record llama.attention.sliding_window. This profile therefore uses the selected GGUF header as serving truth for llama.cpp-style ordinary text decode and records the sliding-window mismatch explicitly." }, "architecture": { "canonical_architecture_id": "solar-pro-preview-instruct-gguf-text", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 22.140032, "swept_params_b": 21.97553664, "auxiliary_resident_params_b": 0.16449536, "resident_weight_gb": 44.282148416, "swept_weight_gb": 43.95239424, "auxiliary_resident_weight_gb": 0.329754176, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for solar-pro-preview-instruct.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 44.282148416 GB. GGUF tensor spans total 44.281384960 GB, while metadata, tokenizer, header, and file overhead account for 0.000763456 GB. Because output.weight is stored separately and the base config records tie_word_embeddings false, token_embd.weight is resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 10, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records llama.context_length 4096, 64 layers, 10 KV heads, 40 attention heads, and 5120 hidden size, implying 128-dimensional key/value heads. Although the base Transformers config has sliding_window 2047 and the Transformers flash/SDPA paths can apply it globally, the selected GGUF header does not include llama.attention.sliding_window, so this selected-artifact profile charges full-context FP16 K/V for llama.cpp-style serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. It does not include the imatrix.dat sidecar unless a separate quantized serving profile explicitly uses it." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the API-selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, kernels, imatrix sidecar handling, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is solar-pro-preview-instruct.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Solar Pro Preview Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/solar-pro-preview-instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff, the API records a public non-gated GGUF text-generation repo with base_model upstage/solar-pro-preview-instruct, quantized tags, region:us, 111877 downloads, GGUF architecture llama, context length 4096, gguf.total 22140032000, and gguf.totalFileSize 44282148416. The API totalFileSize matches solar-pro-preview-instruct.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi Solar Pro Preview Instruct GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/raw/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model upstage/solar-pro-preview-instruct, model_creator upstage, quantized_by MaziyarPanahi, and GGUF/llama.cpp runtime guidance. It lists multiple lower-bit quantized siblings and an imatrix sidecar, but the API-selected totalFileSize points at the FP16 sibling." }, { "label": "MaziyarPanahi Solar Pro GGUF package config stub", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/raw/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff/config.json", "source_type": "config", "supports": [ "package_config_scope" ], "notes": "The package config is only 31 bytes and contains {\"model_type\":\"mistral\"}. The selected GGUF header and base Solar config provide the actual architecture evidence, so this stub is not used for bounds geometry." }, { "label": "Upstage Solar Pro Preview Instruct API metadata", "url": "https://huggingface.co/api/models/upstage/solar-pro-preview-instruct", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "At commit dd4bcf7006df9b1ce3f87711e702e4063832aae3, the base repo is public, MIT licensed, uses custom SolarForCausalLM code, and records BF16 safetensors total 22140032000 parameters." }, { "label": "Upstage Solar Pro Preview Instruct base config", "url": "https://huggingface.co/upstage/solar-pro-preview-instruct/raw/dd4bcf7006df9b1ce3f87711e702e4063832aae3/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The base config records SolarForCausalLM, custom code, BF16, 64 layers, hidden size 5120, intermediate size 17920, 40 attention heads, 10 KV heads, max_position_embeddings 4096, sliding_window 2047, rope_theta 10000, tie_word_embeddings false, and vocab_size 32128." }, { "label": "Upstage Solar custom-code review", "url": "https://huggingface.co/upstage/solar-pro-preview-instruct/raw/dd4bcf7006df9b1ce3f87711e702e4063832aae3/modeling_solar.py", "source_type": "manual_review", "supports": [ "sliding_window_mismatch", "kv_adapter" ], "notes": "The Transformers Solar code imports SlidingWindowCache and applies config.sliding_window in FlashAttention and SDPA causal-mask paths. The vLLM Solar adapter in vllm_solar.py creates regular vLLM Attention layers without an explicit sliding-window layer pattern. Because the selected GGUF header lacks llama.attention.sliding_window metadata, this profile treats the GGUF artifact as full-context serving truth." }, { "label": "MaziyarPanahi Solar Pro GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/tree/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded tree metadata found imatrix.dat 0.012466706 GB plus GGUF siblings: IQ1_S 4.786663232 GB, IQ1_M 5.231488832 GB, IQ2_XS 6.618394432 GB, Q2_K 8.213924672 GB, IQ3_XS 9.125197632 GB, Q3_K_S 9.580672832 GB, Q3_K_M 10.686592832 GB, Q3_K_L 11.635226432 GB, IQ4_XS 11.878032192 GB, Q4_K_S 12.594238272 GB, Q4_K_M 13.310219072 GB, Q5_K_S 15.246070592 GB, Q5_K_M 15.663862592 GB, Q6_K 18.164608832 GB, Q8_0 23.526487872 GB, and FP16 44.282148416 GB. The FP16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi Solar Pro FP16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/resolve/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff/solar-pro-preview-instruct.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 31 metadata entries and 579 tensors. The selected file is 44.282148416 GB, with tensor payloads starting at byte 763456. Tensor spans sum to 44.281384960 GB across 22.140032000B logical elements: token_embd.weight 0.328990720 GB, output.weight 0.328990720 GB, blk.* tensors 43.623383040 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.000763456 GB. Stored tensor bytes split into F16 44.278743040 GB and F32 0.002641920 GB. The header records general.architecture llama, MIT license metadata, llama.block_count 64, context_length 4096, embedding_length 5120, feed_forward_length 17920, attention.head_count 40, attention.head_count_kv 10, rope dimension 128, rope.freq_base 10000, and no llama.attention.sliding_window field." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, package config-stub check, base API/config metadata, custom Solar code review, selected linked-object size checks, and direct selected FP16 GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected Solar Pro Preview Instruct FP16 GGUF artifact. Do not silently substitute lower-bit siblings, the imatrix sidecar, or the base BF16 custom-code checkpoint." }, { "id": "maziyarpanahi--wizardlm-2-7b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/WizardLM-2-7B-GGUF", "title": "MaziyarPanahi WizardLM 2 7B GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of WizardLM 2 7B.", "model_family": "wizardlm-2-7b-mistral-dense-gguf", "base_model_proof": { "base_model": "microsoft/WizardLM-2-7B", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, stale repo-config check, selected linked-object size checks, and base config access check", "config_compatible": false, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of microsoft/WizardLM-2-7B. The package card also says the original model was based on mistralai/Mistral-7B-v0.1. The microsoft base raw config returned HTTP 401 in this audit environment, and the repo-local package config is only a stale model_type mistral stub, so this profile uses the public selected GGUF header as the served architecture source." }, "architecture": { "canonical_architecture_id": "wizardlm-2-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.241732096, "swept_params_b": 7.110660096, "auxiliary_resident_params_b": 0.131072, "resident_weight_gb": 14.484731552, "swept_weight_gb": 14.221852672, "auxiliary_resident_weight_gb": 0.26287888, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for WizardLM-2-7B.fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected F16 linked file is 14.484731552 GB. GGUF tensor payloads total 14.483996672 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000734880 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 32768 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is WizardLM-2-7B.fp16.gguf because its linked object size matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi WizardLM 2 7B GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/WizardLM-2-7B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 491e4644eb4874621cb7d00153a20e0bf7358fb8, the API records a public non-gated Apache-2.0 GGUF repo with base_model microsoft/WizardLM-2-7B, quantized tags, region:us, text-generation pipeline, 112698 downloads, GGUF architecture llama, context length 32768, gguf.total 7241732096, and gguf.totalFileSize 14484731552. The API totalFileSize matches WizardLM-2-7B.fp16.gguf, so this profile targets that artifact." }, { "label": "MaziyarPanahi WizardLM 2 7B GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/raw/491e4644eb4874621cb7d00153a20e0bf7358fb8/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "license", "selected_artifact" ], "notes": "The model card records base_model microsoft/WizardLM-2-7B, model_creator microsoft, quantized_by MaziyarPanahi, Apache-2.0 license metadata, GGUF runtime guidance, and original README text saying WizardLM-2 7B is based on mistralai/Mistral-7B-v0.1." }, { "label": "WizardLM 2 7B base config access check", "url": "https://huggingface.co/microsoft/WizardLM-2-7B/raw/main/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "A raw request for the declared microsoft/WizardLM-2-7B base config returned HTTP 401 with Invalid username or password in this audit environment. This GGUF profile therefore does not claim a direct base-config comparison, but it can use the public selected GGUF header for the served artifact's exact geometry." }, { "label": "MaziyarPanahi WizardLM 2 7B GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/tree/491e4644eb4874621cb7d00153a20e0bf7358fb8", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ3_XS 3.018815200 GB, IQ4_XS 3.944388320 GB, Q2_K 2.719241952 GB, Q3_K_S 3.164567264 GB, Q3_K_M 3.518985952 GB, Q3_K_L 3.822024416 GB, Q4_K_S 4.140373728 GB, Q4_K_M 4.368439008 GB, Q5_K_S 4.997715680 GB, Q5_K_M 5.131409120 GB, Q6_K 5.942064864 GB, Q8_0 7.695857376 GB, and fp16 14.484731552 GB. The F16 size exactly matches API gguf.totalFileSize." }, { "label": "MaziyarPanahi WizardLM 2 7B F16 GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/resolve/491e4644eb4874621cb7d00153a20e0bf7358fb8/WizardLM-2-7B.fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 21 metadata entries and 291 tensors. The selected file is 14.484731552 GB, with tensor payloads starting at byte 734880. Tensor payloads sum to 14.483996672 GB across 7.241732096B logical elements: token_embd.weight 0.262144000 GB, output.weight 0.262144000 GB, blk.* tensors 13.959692288 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.000734880 GB. Stored tensor bytes split into F16 14.482931712 GB and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 32768, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 10000, vocab_size 32000, and no sliding-window field." }, { "label": "MaziyarPanahi WizardLM 2 7B GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/raw/491e4644eb4874621cb7d00153a20e0bf7358fb8/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral. It is recorded as a stale package stub and is not used as the main architecture evidence because the selected GGUF header provides the full served geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, inaccessible base config check, linked GGUF file sizes, stale repo-config check, and direct selected F16 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact API-selected F16 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "maziyarpanahi--yi-1-5-6b-chat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Yi-1.5-6B-Chat-GGUF", "title": "MaziyarPanahi Yi 1.5 6B Chat GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Yi 1.5 6B Chat.", "model_family": "yi-1-5-llama-dense-gguf", "base_model_proof": { "base_model": "01-ai/Yi-1.5-6B-Chat", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, stale repo-config check, and pinned base config comparison", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of 01-ai/Yi-1.5-6B-Chat. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q2_K GGUF header records the same LlamaForCausalLM geometry as the pinned BF16 base config: 32 layers, hidden size 4096, intermediate size 11008, 32 attention heads, 4 KV heads, 128 key/value head dimension, 4096 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "yi-1-5-6b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.06103552, "swept_params_b": 5.79889152, "auxiliary_resident_params_b": 0.262144, "resident_weight_gb": 2.33706688, "swept_weight_gb": 2.249555968, "auxiliary_resident_weight_gb": 0.087510912, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Yi-1.5-6B-Chat.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 2.337066880 GB. GGUF tensor spans total 2.335571968 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.001494912 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 32 Llama decoder layers, 4 KV heads, 128-dimensional key/value heads, and 4096 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. FP16 and other quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.38558871207539136, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is Yi-1.5-6B-Chat.Q2_K.gguf. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Yi 1.5 6B Chat GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 5eeb4c393e2a235fd5dce7f6007dd5ea96577592, the API records a public non-gated GGUF text-generation repo with base_model 01-ai/Yi-1.5-6B-Chat, base_model:quantized metadata, quantized tags, Apache-2.0 tags, region:us, 112573 downloads, GGUF architecture llama, context_length 4096, gguf.total 6061035520, and gguf.totalFileSize 2337066880. The API totalFileSize matches the Q2_K sibling." }, { "label": "MaziyarPanahi Yi 1.5 6B Chat GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/raw/5eeb4c393e2a235fd5dce7f6007dd5ea96577592/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "license" ], "notes": "The card records original model 01-ai/Yi-1.5-6B-Chat, quantized_by MaziyarPanahi, Apache-2.0 license tags, and standard GGUF/llama.cpp runtime guidance." }, { "label": "01-ai Yi 1.5 6B Chat base config and API metadata", "url": "https://huggingface.co/01-ai/Yi-1.5-6B-Chat/raw/771924d1c83d67527d665913415d7086f11ea9c0/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The pinned base config records LlamaForCausalLM, bfloat16, 32 layers, hidden size 4096, intermediate size 11008, 32 attention heads, 4 KV heads, 4096 max positions, 64000 vocabulary size, rope_theta 5000000, and untied embeddings. The base API records Apache-2.0 licensing and safetensors BF16 parameters total 6061035520." }, { "label": "MaziyarPanahi Yi 1.5 6B Chat GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/tree/5eeb4c393e2a235fd5dce7f6007dd5ea96577592", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found Q2_K 2.337066880 GB, Q3_K_S 2.709196672 GB, Q3_K_M 2.992836480 GB, Q3_K_L 3.236892544 GB, Q4_K_S 3.502919552 GB, Q4_K_M 3.673968512 GB, Q5_K_S 4.204154752 GB, Q5_K_M 4.304424832 GB, and Q6_K 4.974284672 GB. The selected Q2_K artifact is the API-selected file for this row." }, { "label": "MaziyarPanahi Yi 1.5 6B Chat Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/resolve/5eeb4c393e2a235fd5dce7f6007dd5ea96577592/Yi-1.5-6B-Chat.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 25 metadata entries and 291 tensors. The linked file is 2.337066880 GB. Tensor spans sum to 2.335571968 GB: output.weight 0.215040000 GB, token_embd.weight 0.086016000 GB, blk.* tensors 2.034499584 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.001494912 GB. Stored tensor spans split into Q2_K 1.231060992 GB, Q3_K 0.850657280 GB, Q4_K 0.037748736 GB, Q6_K 0.215040000 GB, and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 4096, embedding_length 4096, feed_forward_length 11008, attention.head_count 32, attention.head_count_kv 4, rope.dimension_count 128, rope.freq_base 5000000, vocab_size 64000, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Yi 1.5 6B Chat GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/raw/5eeb4c393e2a235fd5dce7f6007dd5ea96577592/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base config and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, linked GGUF file sizes, stale repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected Yi 1.5 6B Chat Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--yi-coder-1-5b-chat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF", "title": "MaziyarPanahi Yi-Coder 1.5B Chat GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the row-selected Q2_K GGUF artifact of Yi-Coder 1.5B Chat.", "model_family": "yi-coder-llama-dense-gguf", "base_model_proof": { "base_model": "01-ai/Yi-Coder-1.5B-Chat", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, stale repo-config check, and pinned base config comparison", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of 01-ai/Yi-Coder-1.5B-Chat. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q2_K GGUF header records the same LlamaForCausalLM geometry as the pinned BF16 base config: 24 layers, hidden size 2048, intermediate size 5504, 16 attention heads, 16 KV heads, 128 key/value head dimension, 131072 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "yi-coder-1-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.47649536, "swept_params_b": 1.34542336, "auxiliary_resident_params_b": 0.131072, "resident_weight_gb": 0.634699968, "swept_weight_gb": 0.590200832, "auxiliary_resident_weight_gb": 0.044499136, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Yi-Coder-1.5B-Chat.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 0.634699968 GB. GGUF tensor spans total 0.633208832 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.001491136 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q6_K, IQ4_NL, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 24 Llama decoder layers, 16 KV heads, 128-dimensional key/value heads, and 131072 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the row-selected Q2_K GGUF artifact. FP16 and other quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.42986926013773585, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets Yi-Coder-1.5B-Chat.Q2_K.gguf because the catalog row is the 2-bit GGUF entry. The live HF API gguf.totalFileSize currently matches the FP16 sibling instead, so exact selected-file header bytes are authoritative. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Yi-Coder 1.5B Chat GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit cc8525f15600e412829215092967f636b654913a, the API records a public non-gated GGUF text-generation repo with base_model 01-ai/Yi-Coder-1.5B-Chat, base_model:quantized metadata, quantized tags, region:us, 124754 downloads, GGUF architecture llama, context_length 131072, gguf.total 1476495360, and gguf.totalFileSize 2954682336. The API totalFileSize matches the fp16 sibling, while this profile targets the Q2_K file for the catalog's 2-bit row." }, { "label": "MaziyarPanahi Yi-Coder 1.5B Chat GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model 01-ai/Yi-Coder-1.5B-Chat, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base-model API record Apache-2.0 licensing." }, { "label": "01-ai Yi-Coder 1.5B Chat base config and API metadata", "url": "https://huggingface.co/01-ai/Yi-Coder-1.5B-Chat/raw/92fdd1b2f1539ac990e7f4a921db5601da2f0299/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The pinned base config records LlamaForCausalLM, bfloat16, 24 layers, hidden size 2048, intermediate size 5504, 16 attention heads, 16 KV heads, 131072 max positions, 64000 vocabulary size, and untied embeddings. The base API records Apache-2.0 licensing and safetensors BF16 parameters total 1476495360." }, { "label": "MaziyarPanahi Yi-Coder 1.5B Chat GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF/tree/cc8525f15600e412829215092967f636b654913a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ1_S 0.491167936 GB, IQ1_M 0.508567744 GB, IQ2_XS 0.563912896 GB, Q2_K 0.634699968 GB, IQ3_XS 0.694952128 GB, Q3_K_S 0.723411136 GB, Q3_K_M 0.785719488 GB, Q3_K_L 0.826040512 GB, IQ4_XS 0.832569536 GB, Q4_K_S 0.904184000 GB, Q4_K_M 0.963674304 GB, Q5_K_S 1.051230400 GB, Q5_K_M 1.100185792 GB, and fp16 2.954682336 GB. The selected Q2_K artifact is the standard 2-bit GGUF file for this row." }, { "label": "MaziyarPanahi Yi-Coder 1.5B Chat Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF/resolve/cc8525f15600e412829215092967f636b654913a/Yi-Coder-1.5B-Chat.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 219 tensors. The linked file is 0.634699968 GB. Tensor spans sum to 0.633208832 GB: output.weight 0.107520000 GB, token_embd.weight 0.043008000 GB, blk.* tensors 0.482672640 GB, and output_norm.weight 0.000008192 GB. Metadata/tokenizer/header/file overhead accounts for 0.001491136 GB. Stored tensor spans split into Q2_K 0.286605312 GB, Q3_K 0.086507520 GB, Q6_K 0.107520000 GB, IQ4_NL 0.152174592 GB, and F32 0.000401408 GB. The header records general.architecture llama, Apache-2.0 license, llama.block_count 24, context_length 131072, embedding_length 2048, feed_forward_length 5504, attention.head_count 16, attention.head_count_kv 16, rope.dimension_count 128, rope.freq_base 10000000, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Yi-Coder 1.5B Chat GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF/raw/cc8525f15600e412829215092967f636b654913a/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base config and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, linked GGUF file sizes, stale repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the row-selected Yi-Coder 1.5B Chat Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "maziyarpanahi--yi-coder-9b-chat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MaziyarPanahi/Yi-Coder-9B-Chat-GGUF", "title": "MaziyarPanahi Yi-Coder 9B Chat GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the row-selected Q2_K GGUF artifact of Yi-Coder 9B Chat.", "model_family": "yi-coder-llama-dense-gguf", "base_model_proof": { "base_model": "01-ai/Yi-Coder-9B-Chat", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q2_K GGUF header metadata, stale repo-config check, and pinned base config comparison", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of 01-ai/Yi-Coder-9B-Chat. The repo-local config.json incorrectly says model_type mistral, so this profile does not use it as architecture evidence. The selected Q2_K GGUF header records the same LlamaForCausalLM geometry as the pinned BF16 base config: 48 layers, hidden size 4096, intermediate size 11008, 32 attention heads, 4 KV heads, 128 key/value head dimension, 131072 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "yi-coder-9b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.829407232, "swept_params_b": 8.567263232, "auxiliary_resident_params_b": 0.262144, "resident_weight_gb": 3.354325792, "swept_weight_gb": 3.26680576, "auxiliary_resident_weight_gb": 0.087520032, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Yi-Coder-9B-Chat.Q2_K.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q2_K linked file is 3.354325792 GB. GGUF tensor spans total 3.352821760 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.001504032 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 48 Llama decoder layers, 4 KV heads, 128-dimensional key/value heads, and 131072 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the row-selected Q2_K GGUF artifact. FP16 and other quantized siblings should get separate profiles if selected." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3799038490197934, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "This profile targets Yi-Coder-9B-Chat.Q2_K.gguf because the catalog row is the 2-bit GGUF entry. The live HF API gguf.totalFileSize currently matches the FP16 sibling instead, so exact selected-file header bytes are authoritative. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "MaziyarPanahi Yi-Coder 9B Chat GGUF API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit a1ceaae2d11b41e263560af6f806b9aedbefffd2, the API records a public non-gated GGUF text-generation repo with base_model 01-ai/Yi-Coder-9B-Chat, base_model:quantized metadata, quantized tags, region:us, 118204 downloads, GGUF architecture llama, context_length 131072, gguf.total 8829407232, and gguf.totalFileSize 17661112896. The API totalFileSize matches the fp16 sibling, while this profile targets the Q2_K file for the catalog's 2-bit row." }, { "label": "MaziyarPanahi Yi-Coder 9B Chat GGUF model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records original model 01-ai/Yi-Coder-9B-Chat, quantized_by MaziyarPanahi, and standard GGUF/llama.cpp runtime guidance. It does not declare a license in cardData, but the selected GGUF metadata and base-model API record Apache-2.0 licensing." }, { "label": "01-ai Yi-Coder 9B Chat base config and API metadata", "url": "https://huggingface.co/01-ai/Yi-Coder-9B-Chat/raw/356a1f8d4e4a606d0b879e54191ca809918576b8/config.json", "source_type": "config", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter", "license" ], "notes": "The pinned base config records LlamaForCausalLM, bfloat16, 48 layers, hidden size 4096, intermediate size 11008, 32 attention heads, 4 KV heads, 131072 max positions, 64000 vocabulary size, rope_theta 10000000, and untied embeddings. The base API records Apache-2.0 licensing." }, { "label": "MaziyarPanahi Yi-Coder 9B Chat GGUF linked-object size checks", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF/tree/a1ceaae2d11b41e263560af6f806b9aedbefffd2", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found IQ1_S 2.014573344 GB, IQ1_M 2.181640992 GB, IQ2_XS 2.708009760 GB, Q2_K 3.354325792 GB, Q3_K_S 3.899208480 GB, Q3_K_M 4.324406048 GB, Q3_K_L 4.690752288 GB, IQ4_XS 4.785009440 GB, Q4_K_S 5.071860512 GB, Q4_K_M 5.328958240 GB, Q5_K_S 6.107853600 GB, Q5_K_M 6.258258720 GB, and fp16 17.661112896 GB. The selected Q2_K artifact is the standard 2-bit GGUF file for this row." }, { "label": "MaziyarPanahi Yi-Coder 9B Chat Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF/resolve/a1ceaae2d11b41e263560af6f806b9aedbefffd2/Yi-Coder-9B-Chat.Q2_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 35 metadata entries and 435 tensors. The linked file is 3.354325792 GB. Tensor spans sum to 3.352821760 GB: output.weight 0.215040000 GB, token_embd.weight 0.086016000 GB, blk.* tensors 3.051749376 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.001504032 GB. Stored tensor spans split into Q2_K 1.803583488 GB, Q3_K 1.275985920 GB, Q4_K 0.056623104 GB, Q6_K 0.215040000 GB, and F32 0.001589248 GB. The header records general.architecture llama, Apache-2.0 license, llama.block_count 48, context_length 131072, embedding_length 4096, feed_forward_length 11008, attention.head_count 32, attention.head_count_kv 4, rope.dimension_count 128, rope.freq_base 10000000, and a separate output.weight tensor." }, { "label": "MaziyarPanahi Yi-Coder 9B Chat GGUF stale config", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF/raw/a1ceaae2d11b41e263560af6f806b9aedbefffd2/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral, conflicting with the base config and selected GGUF header. It is recorded as stale and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned base config/API metadata, linked GGUF file sizes, stale repo-config check, and direct selected Q2_K GGUF tensor-index range read." }, "notes": "Use this profile for the row-selected Yi-Coder 9B Chat Q2_K GGUF artifact. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "meta-llama--llama-2-13b-chat-hf", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-2-13b-chat-hf", "title": "Meta Llama 2 13B Chat HF F16", "summary": "Unsupported profile stub for the gated Llama 2 13B Chat HF repo.", "model_family": "llama2-dense", "architecture": { "canonical_architecture_id": "llama-2-13b-chat", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 13.01586688, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 13015864320 F16 and 2560 F32 safetensors parameters for this repo. KV geometry, max context, tied embeddings, and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval. Raw config and model.safetensors.index.json HEAD checks returned 401 GatedRepo responses, and hf download config.json and model.safetensors.index.json returned access denied in this audit environment.", "notes": "Do not infer Llama 2 layer count, attention head count, KV heads, head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.0000003933660393, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "F16 plus tiny F32 weight dtype split comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 2 13B Chat HF API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-13b-chat-hf", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit a2cb7a712bb6e5e736ca7f8cd98167f81a0b5bd8, the API reports gated: manual, text-generation pipeline, Transformers library, Llama 2 license, region:us, 115324 downloads, and safetensors counts F16 13015864320 and F32 2560." }, { "label": "Gated config and tensor-index access checks", "url": "https://huggingface.co/meta-llama/Llama-2-13b-chat-hf/raw/a2cb7a712bb6e5e736ca7f8cd98167f81a0b5bd8/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated pinned config and model.safetensors.index.json HEAD requests returned 401 GatedRepo responses. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-2-7b-chat-hf", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-2-7b-chat-hf", "title": "Meta Llama 2 7B Chat HF F16", "summary": "Unsupported profile stub for the gated Llama 2 7B Chat HF repo.", "model_family": "llama2-dense", "architecture": { "canonical_architecture_id": "llama-2-7b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 6.738417664, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 6738417664 F16 safetensors parameters for this repo. KV geometry and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval. Raw config, generation config, tokenizer config, model.safetensors.index.json, and authenticated HF CLI downloads returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "F16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 2 7B Chat HF API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-7b-chat-hf", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit f5db02db724555f92da89c216ac04704f23d4590, the API reports gated: manual, text-generation pipeline, Transformers library, Llama 2 license, region:us, 265563 downloads, and safetensors count F16 6738417664." }, { "label": "Raw file access checks", "url": "https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/raw/f5db02db724555f92da89c216ac04704f23d4590/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Authenticated raw requests for config.json, generation_config.json, tokenizer_config.json, and model.safetensors.index.json returned restricted-access 401 responses in this environment." }, { "label": "HF CLI authenticated access check", "url": "https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/resolve/f5db02db724555f92da89c216ac04704f23d4590/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "HF CLI was authenticated as osolmaz, but hf download for config.json and model.safetensors.index.json returned 403 Forbidden: this repository requires approval and the user is not in the authorized list." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-2-7b-hf", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-2-7b-hf", "title": "Meta Llama 2 7B HF F16", "summary": "Unsupported profile stub for the gated Llama 2 7B HF repo.", "model_family": "llama2-dense", "architecture": { "canonical_architecture_id": "llama-2-7b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 6.738417664, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 6738415616 F16 and 2048 F32 safetensors parameters for this repo. KV geometry and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, model.safetensors.index.json, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.000000607857839, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "F16 plus tiny F32 weight dtype split comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 2 7B HF API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-7b-hf", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 01c7f73d771dfac7d292323805ebc428287df4f9, the API reports gated: manual, text-generation pipeline, Llama 2 license, region:us, 679555 downloads, and safetensors counts F16 6738415616 and F32 2048." }, { "label": "Gated config, card, and tensor-index access check", "url": "https://huggingface.co/meta-llama/Llama-2-7b-hf/raw/01c7f73d771dfac7d292323805ebc428287df4f9/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 restricted-access responses. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-1-405b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.1-405B-Instruct", "title": "Meta Llama 3.1 405B Instruct BF16", "summary": "Unsupported profile stub for the gated Llama 3.1 405B Instruct repo.", "model_family": "llama31-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-405B", "relation": "finetune", "source": "Hugging Face API base_model metadata", "config_compatible": false, "notes": "The public API metadata identifies meta-llama/Llama-3.1-405B as the finetune base, but the gated config and tensor index are not accessible in this environment, so architecture compatibility is not audited." }, "architecture": { "canonical_architecture_id": "llama-3-1-405b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 405.8533888, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 405853388800 BF16 safetensors parameters for this repo. KV geometry, max context, tied embeddings, and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, model.safetensors.index.json, a safetensors shard HEAD request, and hf download checks returned access denied in this audit environment.", "notes": "Do not infer Llama 3.1 layer count, attention head count, KV heads, head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.1 405B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-405B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit be673f326cab4cd22ccfef76109faf68e41aa5f1, the API reports gated: manual, text-generation pipeline, Llama 3.1 license, current downloads 210625, base_model:meta-llama/Llama-3.1-405B, endpoints_compatible, region:us, 191 safetensors shards, and safetensors count BF16 405853388800." }, { "label": "Gated config and tensor access checks", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct/raw/be673f326cab4cd22ccfef76109faf68e41aa5f1/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw HEAD checks for config.json, model.safetensors.index.json, and model-00001-of-00191.safetensors returned 401 GatedRepo responses. The README was publicly reachable, but it does not provide the tensor-header evidence needed for production bounds. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-1-405b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.1-405B", "title": "Meta Llama 3.1 405B BF16", "summary": "Unsupported profile stub for the gated Llama 3.1 405B base repo.", "model_family": "llama31-dense", "architecture": { "canonical_architecture_id": "llama-3-1-405b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 405.8533888, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 405853388800 BF16 safetensors parameters for this repo. KV geometry, max context, tied embeddings, and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, model.safetensors.index.json, a safetensors shard HEAD request, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama 3.1 layer count, attention head count, KV heads, head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.1 405B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-405B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit b906e4dc842aa489c962f9db26554dcfdde901fe, the API reports gated: manual, text-generation pipeline, Llama 3.1 license, current downloads 230262, endpoints_compatible, region:us, 191 safetensors shards, and safetensors count BF16 405853388800." }, { "label": "Gated config, card, and tensor access checks", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B/raw/b906e4dc842aa489c962f9db26554dcfdde901fe/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 GatedRepo responses. A HEAD request for model-00001-of-00191.safetensors also returned 401 GatedRepo. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-1-70b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.1-70B-Instruct", "title": "Meta Llama 3.1 70B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.1 70B Instruct repo.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-70B", "relation": "finetune", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API identifies a base model, but the gated config was not accessible to this audit environment, so architecture compatibility could not be verified." }, "architecture": { "canonical_architecture_id": "llama-3-1-70b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 70.553706496, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 70553706496 BF16 safetensors parameters for this repo. KV geometry and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, model.safetensors.index.json, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.1 70B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-70B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 1605565b47bb9346c5515c34102e054115b4f98b, the API reports gated: manual, text-generation pipeline, license llama3.1, region:us, 1121582 downloads, and BF16 safetensors count 70553706496." }, { "label": "Gated config, card, and tensor-index access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct/raw/1605565b47bb9346c5515c34102e054115b4f98b/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 restricted-access responses. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-1-8b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.1-8B-Instruct", "title": "Meta Llama 3.1 8B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.1 8B Instruct repo.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B", "relation": "finetune", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API identifies a base model, but the gated config was not accessible to this audit environment, so architecture compatibility could not be verified." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.030261248, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8030261248 BF16 safetensors parameters for this repo. KV geometry is not audited because the config is gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and both raw config access and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, context length, or RoPE settings from the model name. Replace this with an audited full_context adapter only after config access is available." }, "notes": "This profile intentionally fails closed until the gated config can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.1 8B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-8B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving" ], "notes": "The API reports gated: manual, text-generation pipeline, license llama3.1, BF16 safetensors count 8030261248, and base-model metadata." }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config returned 401, and hf download with the configured CLI token returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config is not accessible in this audit environment, so KV geometry and max context cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config evidence is available." }, { "id": "meta-llama--llama-3-1-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.1-8B", "title": "Meta Llama 3.1 8B BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.1 8B base repo.", "model_family": "llama3.1-dense", "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.030261248, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8030261248 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw weights, and hf download for config.json plus model.safetensors.index.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.1 8B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-8B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit d04e592bb4f6aa9cfee91e2e20afa771667e1d4b, the API reports gated: manual, text-generation pipeline, license llama3.1, 1,517,664 downloads, and BF16 safetensors count 8030261248." }, { "label": "Gated config and weight access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B/raw/d04e592bb4f6aa9cfee91e2e20afa771667e1d4b/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 GatedRepo. hf download with the configured CLI token for osolmaz returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-2-1b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.2-1B-Instruct", "title": "Meta Llama 3.2 1B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.2 1B Instruct repo.", "model_family": "llama3.2-dense", "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 1.2358144, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 1235814400 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, safetensors index, and hf download config.json all returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, context length, or RoPE settings from the model name. Replace this with an audited adapter only after direct config evidence is available." }, "notes": "This profile intentionally fails closed until the gated config can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.2 1B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "The API reports gated: manual, text-generation pipeline, license llama3.2, and BF16 safetensors count 1235814400." }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config returned 401, the safetensors index returned 401, and hf download with the configured CLI token returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config is not accessible in this audit environment, so KV geometry and max context cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config evidence is available." }, { "id": "meta-llama--llama-3-2-1b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.2-1B", "title": "Meta Llama 3.2 1B BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.2 1B base repo.", "model_family": "llama3.2-dense", "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 1.2358144, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 1235814400 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors bytes are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw weights, and hf download for config.json plus model.safetensors returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.2 1B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 4e20de362430cd3b72f300e6b0f18e50e7166e08, the API reports gated: manual, text-generation pipeline, license llama3.2, and BF16 safetensors count 1235814400." }, { "label": "Meta Llama 3.2 1B public model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B", "source_type": "model_card", "supports": [ "max_context_tokens", "architecture" ], "notes": "The public model card reports Llama 3.2 text-only 1B as 1.23B parameters with 128k context, GQA, and shared embeddings. That is sufficient for catalog display, but not enough to audit exact KV geometry and swept tensor traffic." }, { "label": "Gated config and weight access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B/raw/4e20de362430cd3b72f300e6b0f18e50e7166e08/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, generation config, README, and model.safetensors requests returned 401. hf download with the configured CLI token for osolmaz returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact max context encoding, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-2-3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.2-3B-Instruct", "title": "Meta Llama 3.2 3B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.2 3B Instruct repo.", "model_family": "llama3.2-dense", "architecture": { "canonical_architecture_id": "llama-3-2-3b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 3.212749824, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 3212749824 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw safetensors index, and hf download for config.json plus model.safetensors.index.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, context length, or RoPE settings from the model name. Replace this with an audited adapter only after direct config evidence is available." }, "notes": "This profile intentionally fails closed until the gated config can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.2 3B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 0cb88a4f764b7a12671c53f0838cd831a0843b95, the API reports gated: manual, text-generation pipeline, license llama3.2, and BF16 safetensors count 3212749824." }, { "label": "Gated config and index access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/raw/0cb88a4f764b7a12671c53f0838cd831a0843b95/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config and safetensors index requests returned 401. hf download with the configured CLI token for osolmaz returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor index are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-2-3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.2-3B", "title": "Meta Llama 3.2 3B BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.2 3B base repo.", "model_family": "llama3.2-dense", "architecture": { "canonical_architecture_id": "llama-3-2-3b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 3.212749824, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 3212749824 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, safetensors index, model weights, and hf download for config.json plus model.safetensors.index.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, context length, RoPE settings, tied embeddings, or swept decode traffic from the model name or nearby instruct repo. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.2 3B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 13afe5124825b4f3751f836b40dafda64c1ed062, the API reports gated: manual, text-generation pipeline, library transformers, license llama3.2, region:us, current downloads 640543, and BF16 safetensors count 3212749824." }, { "label": "Gated config, card, index, and weight access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B/raw/13afe5124825b4f3751f836b40dafda64c1ed062/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config.json, README.md, model.safetensors.index.json, and model.safetensors requests returned 401 restricted-access responses. hf download with the configured CLI token for osolmaz returned access denied because the repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context settings, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-3-3-70b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-3.3-70B-Instruct", "title": "Meta Llama 3.3 70B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 3.3 70B Instruct repo.", "model_family": "llama3.3-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-70B", "relation": "finetune", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API identifies a base model, but the gated config was not accessible to this audit environment, so architecture compatibility could not be verified." }, "architecture": { "canonical_architecture_id": "llama-3-3-70b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 70.553706496, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 70553706496 BF16 safetensors parameters for this repo. KV geometry and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, model.safetensors.index.json, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3.3 70B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.3-70B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b, the API reports gated: manual, text-generation pipeline, library transformers, license llama3.3, region:us, 740835 downloads, base_model metadata for meta-llama/Llama-3.1-70B, and BF16 safetensors count 70553706496." }, { "label": "Gated config, card, and tensor-index access check", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/raw/6f6073b423013f6a7d4d9f39144961bfbfbc386b/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 restricted-access responses. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-4-maverick-17b-128e-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "title": "Meta Llama 4 Maverick 17B 128E Instruct FP8", "summary": "Unsupported profile stub for the gated FP8 Llama 4 Maverick 17B 128E Instruct repo.", "model_family": "llama4-maverick-multimodal-moe", "base_model_proof": { "base_model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "relation": "quantized", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API self-references this repo as both base_model and base_model:quantized. That metadata is preserved for traceability, but it is not usable base-model proof because the served config and tensor index are gated." }, "architecture": { "canonical_architecture_id": "llama-4-maverick-17b-128e-fp8", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 401.649841664, "parameter_scope": "hf_api_safetensors_total_only_gated_tensor_layout_unverified", "notes": "The Hugging Face API records 401649841664 safetensors parameters: 15102785024 BF16 parameters and 386547056640 F8_E4M3 parameters. The shallow API config exposes Llama4ForConditionalGeneration/model_type llama4 and compressed-tensors 8-bit Linear quantization, but raw config, card, and tensor index files are gated. This API total is preserved only as package metadata; resident, active, routed-expert, multimodal, and swept text-decode traffic are not audited." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, generation_config.json, recipe.yaml, model.safetensors.index.json, and hf download config.json are inaccessible with the configured osolmaz CLI identity.", "notes": "Do not infer Llama 4 Maverick MoE routing, multimodal components, KV heads, context length, tied embeddings, or swept decode traffic from the repo name or from the ungated API parameter total. Replace this with an audited adapter only after direct gated config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.037601869731681, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "The API identifies compressed-tensors FP8 quantization with BF16 side tensors, but production bounds are disabled because KV geometry, MoE routing, and swept tensor traffic are unavailable.", "notes": "The effective 1.037601869731681 bytes/parameter value is computed from API safetensors dtype counts only: BF16 parameters charged at 2 bytes and F8_E4M3 parameters charged at 1 byte. It is not used for production throughput because status is unsupported." }, "evidence": [ { "label": "Meta Llama 4 Maverick FP8 API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 94125d2bd83076b21eed33119525e29eaf3894f4, the API reports gated: manual, library_name transformers, image-text-to-text pipeline, architecture Llama4ForConditionalGeneration, model_type llama4, compressed-tensors 8-bit Linear quantization, license other / llama4, region:us, 108194 downloads, and safetensors counts BF16 15102785024, F8_E4M3 386547056640, total 401649841664." }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8/raw/94125d2bd83076b21eed33119525e29eaf3894f4/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw pinned config.json returned HTTP 401 GatedRepo. hf download config.json with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." }, { "label": "Gated card, generation config, recipe, and tensor-index access checks", "url": "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8/raw/94125d2bd83076b21eed33119525e29eaf3894f4/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw pinned README.md, generation_config.json, recipe.yaml, and model.safetensors.index.json requests returned HTTP 401 GatedRepo responses. Direct tensor grouping, MoE expert layout, full context settings, and KV/cache geometry therefore cannot be audited." } ], "unsupported_reason": "Gated raw config, model card, generation config, recipe, and tensor index files are not accessible in this audit environment, so multimodal architecture, MoE routing, KV geometry, max context, active/swept traffic, resident component split, and exact FP8/BF16 tensor grouping cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after accepted-gate access allows direct config and safetensors header evidence." }, { "id": "meta-llama--llama-4-scout-17b-16e-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "title": "Meta Llama 4 Scout 17B 16E Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Llama 4 Scout 17B 16E Instruct repo.", "model_family": "llama4-scout-multimodal-moe", "base_model_proof": { "base_model": "meta-llama/Llama-4-Scout-17B-16E", "relation": "finetune", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API identifies a base model, but the gated config was not accessible to this audit environment, so architecture compatibility could not be verified." }, "architecture": { "canonical_architecture_id": "llama-4-scout-17b-16e", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 108.641793536, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 108641793536 BF16 safetensors parameters for this repo. This is the only directly accessible parameter evidence; active MoE traffic, multimodal residency, and swept decode traffic are not audited because the config and tensor index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires manual approval, and raw config, README, model.safetensors.index.json, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama 4 MoE routing, multimodal components, KV heads, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry, MoE routing, and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 4 Scout 17B 16E Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-4-Scout-17B-16E-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 92f3b1597a195b523d8d9e5700e57e4fbb8f20d3, the API reports gated: manual, image-text-to-text pipeline, license other, region:us, 725196 downloads, and BF16 safetensors count 108641793536." }, { "label": "Gated config, card, and tensor-index access check", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/raw/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401 restricted-access responses. hf download with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so multimodal architecture, MoE routing, KV geometry, max context, active/swept traffic, and resident component split cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--llama-guard-3-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-Guard-3-8B", "title": "Meta Llama Guard 3 8B BF16", "summary": "Unsupported profile stub for the gated BF16 Llama Guard 3 8B repo.", "model_family": "llama3.1-guard-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B", "relation": "finetune", "source": "Hugging Face API model metadata", "config_compatible": false, "notes": "The public API identifies meta-llama/Llama-3.1-8B as the base model, but the gated Llama Guard config and tensor index are not accessible to this audit environment, so architecture compatibility and safety-head/tokenizer behavior cannot be directly verified." }, "architecture": { "canonical_architecture_id": "llama-guard-3-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.030261248, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8030261248 BF16 safetensors parameters for this repo. Resident and swept text-decode traffic are not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, README, model.safetensors.index.json, original/params.json, and hf download config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, context length, RoPE settings, tied embeddings, or swept decode traffic from the model name or from the public API parameter total. Replace this with an audited adapter only after accepted-gate access allows direct config and tensor-header evidence." }, "notes": "This profile intentionally fails closed until the gated config and tensor bytes can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama Guard 3 8B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-Guard-3-8B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 7327bd9f6efbbe6101dc6cc4736302b3cbb6e425, the API reports gated: manual, text-generation pipeline, transformers library, Llama 3.1 license, region:us, 244402 downloads, base_model metadata for meta-llama/Llama-3.1-8B, and BF16 safetensors count 8030261248." }, { "label": "Gated raw file access check", "url": "https://huggingface.co/meta-llama/Llama-Guard-3-8B/raw/7327bd9f6efbbe6101dc6cc4736302b3cbb6e425/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config.json, README.md, model.safetensors.index.json, and original/params.json requests all returned 401 GatedRepo responses. `hf download meta-llama/Llama-Guard-3-8B config.json --revision 7327bd9f6efbbe6101dc6cc4736302b3cbb6e425` with the configured osolmaz CLI identity returned access denied because the repository requires approval." }, { "label": "Weni Llama Guard 3 8B AWQ audited sibling", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "A public AWQ derivative is audited separately from its own served config and tensor headers, but it does not make the gated Meta BF16 base executable config or tensor layout directly audited. This fail-closed base profile deliberately avoids copying geometry from the derivative." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, max context, tied embeddings, safety-model tokenizer/config details, resident/swept traffic, and exact tensor layout cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after accepted-gate access allows direct config, tokenizer, and safetensors header evidence." }, { "id": "meta-llama--llama-guard-4-12b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Llama-Guard-4-12B", "title": "Meta Llama Guard 4 12B BF16", "summary": "Unsupported profile stub for the manually gated BF16 Llama Guard 4 12B safety repo.", "model_family": "llama-guard-4-multimodal-safety", "architecture": { "canonical_architecture_id": "llama-guard-4-12b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 12.001097216, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 12001097216 BF16 safetensors parameters for this repo. Text layer geometry, multimodal adapter geometry, context length, tied embedding behavior, and swept ordinary decode traffic are not audited because the gated config and safetensors headers are inaccessible to the configured HF token." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is manually gated. Anonymous raw config, README, and safetensors index requests return 401, and authenticated requests with the configured osolmaz HF CLI identity return 403 access denied.", "notes": "Do not infer Llama 4 layer count, KV heads, head dimension, context length, image-token behavior, multimodal projector residency, tied embedding behavior, or swept decode traffic from the model name or API parameter total. Replace this with an audited adapter only after direct config and tensor-header evidence is available." }, "notes": "This profile intentionally fails closed until the gated Llama Guard 4 config and tensor layout can be audited directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry, multimodal residency, and swept tensor traffic cannot be verified.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama Guard 4 12B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-Guard-4-12B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "license", "pipeline", "unsupported_reason" ], "notes": "At commit 87acb4b94e930c3d679e6e7ee9d57e2feab9ea71, the API reports gated: manual, image-text-to-text pipeline, license:other, safety and llama4 tags, region:us, 153621 downloads, and BF16 safetensors count 12001097216." }, { "label": "Meta Llama Guard 4 12B gated config and tensor-index access checks", "url": "https://huggingface.co/meta-llama/Llama-Guard-4-12B/raw/87acb4b94e930c3d679e6e7ee9d57e2feab9ea71/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Anonymous raw requests for config.json, README.md, and model.safetensors.index.json returned HTTP 401 access-restricted responses. With the configured HF CLI identity, authenticated raw requests for the same files returned HTTP 403 access denied. The public API exposes the BF16 safetensors parameter total and file names, but not layer count, KV heads, head dimension, context length, multimodal tensor grouping, tied embeddings, or tensor byte layout." } ], "unsupported_reason": "The repo's config and tensor headers are gated and inaccessible to the configured audit identity, so KV geometry, max context, multimodal residency, tied embeddings, and swept ordinary text-decode traffic cannot be verified without guessing.", "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-07", "notes": "Reviewed from current HF API metadata plus anonymous and authenticated raw config/README/safetensors-index access checks." }, "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct Llama Guard 4 config and safetensors header evidence is available." }, { "id": "meta-llama--meta-llama-3-70b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Meta-Llama-3-70B", "title": "Meta Llama 3 70B BF16", "summary": "Unsupported profile stub for the gated BF16 Meta Llama 3 70B repo.", "model_family": "llama3-dense", "architecture": { "canonical_architecture_id": "llama-3-70b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 70.553706496, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 70553706496 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw README, and raw safetensors index returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3 70B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-70B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit c82494877ce7f6d7d317c56ec081328e382c72fe, the live API reports gated: manual, text-generation pipeline, license llama3, 149077 downloads, region:us, and BF16 safetensors count 70553706496." }, { "label": "Meta Llama 3 70B gated config and index access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-70B/raw/c82494877ce7f6d7d317c56ec081328e382c72fe/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config.json, README.md, and model.safetensors.index.json requests returned HTTP 401 access-restricted responses in this audit environment." } ], "unsupported_reason": "Gated config and tensor index are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from current HF API metadata and direct raw config/README/safetensors-index access checks." }, "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--meta-llama-3-8b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Meta-Llama-3-8B-Instruct", "title": "Meta Llama 3 8B Instruct BF16", "summary": "Unsupported profile stub for the gated BF16 Meta Llama 3 8B Instruct repo.", "model_family": "llama3-dense", "architecture": { "canonical_architecture_id": "llama-3-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.030261248, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8030261248 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw safetensors index, raw README, and hf download for config.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3 8B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-8B-Instruct", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 8afb486c1db24fe5011ec46dfbe5b5dccdb575c2, the API reports gated: manual, text-generation pipeline, license llama3, 1,306,206 downloads, region:us, and BF16 safetensors count 8030261248." }, { "label": "Gated config and index access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/raw/8afb486c1db24fe5011ec46dfbe5b5dccdb575c2/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401. hf download with the configured CLI token returned 403 access denied because the repository requires approval and the account is not in the authorized list." } ], "unsupported_reason": "Gated config and tensor index are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "meta-llama--meta-llama-3-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "meta-llama/Meta-Llama-3-8B", "title": "Meta Llama 3 8B BF16", "summary": "Unsupported profile stub for the gated BF16 Meta Llama 3 8B base repo.", "model_family": "llama3-dense", "architecture": { "canonical_architecture_id": "llama-3-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.030261248, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8030261248 BF16 safetensors parameters for this repo. KV geometry is not audited because the config and safetensors index are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo requires approval, and raw config, raw safetensors index, raw README, and hf download for config.json and model.safetensors.index.json returned access denied in this audit environment.", "notes": "Do not infer Llama KV heads, exact head dimension, RoPE settings, context length, tied embeddings, or swept decode traffic from the model name. Replace this with an audited adapter only after direct config and tensor evidence is available." }, "notes": "This profile intentionally fails closed until the gated config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry is unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "Meta Llama 3 8B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-8B", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 8cde5ca8380496c9a6cc7ef3a8b46a0372a1d920, the API reports gated: manual, text-generation pipeline, license llama3, 1,237,078 downloads, region:us, and BF16 safetensors count 8030261248." }, { "label": "Gated config and index access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B/raw/8cde5ca8380496c9a6cc7ef3a8b46a0372a1d920/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config, README, and model.safetensors.index.json requests returned 401. hf download with the configured CLI token returned access denied for both config.json and model.safetensors.index.json because the repository requires approval and the account is not in the authorized list." } ], "unsupported_reason": "Gated config and tensor index are not accessible in this audit environment, so KV geometry, max context, tied embeddings, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct config and safetensors header evidence is available." }, { "id": "microsoft--biogpt", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/biogpt", "title": "Microsoft BioGPT FP32", "summary": "Audited memory-side text-decode bounds profile for the FP32 PyTorch BioGPT repo.", "model_family": "biogpt-dense", "base_model_proof": { "base_model": "microsoft/biogpt", "relation": "base", "source": "Hugging Face model metadata, pinned BioGPT config and model card, range-read PyTorch checkpoint metadata, and Transformers BioGPT implementation review", "config_compatible": true, "notes": "BioGPT is the target repo itself. The config records BioGptForCausalLM with decoder-only BioGPT geometry, and the pinned PyTorch checkpoint metadata matches that architecture." }, "architecture": { "canonical_architecture_id": "biogpt", "max_context_tokens": 1024, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.346763264, "swept_params_b": 0.34571264, "auxiliary_resident_params_b": 0.001050624, "resident_weight_gb": 1.387053056, "swept_weight_gb": 1.38285056, "auxiliary_resident_weight_gb": 0.004202496, "resident_parameter_scope": "unique FP32 parameters after Transformers BioGPT output-projection weight tying", "swept_parameter_scope": "ordinary text decode sweeps biogpt.layers.*, biogpt.layer_norm.*, and the tied biogpt.embed_tokens/output_projection output matrix", "auxiliary_scope": "biogpt.embed_positions.weight has 1026 x 1024 FP32 parameters for learned position lookup and is resident-only for ordinary generated tokens", "notes": "The raw PyTorch zip stores 389 FloatStorage tensors, including a duplicate output_projection.weight. Transformers declares output_projection.weight tied to biogpt.embed_tokens.weight, so production resident bytes charge the unique tied serving graph rather than the serialized duplicate." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 24 decoder layers, 16 attention heads, hidden size 1024, and BioGPT attention with separate K/V projections cached through DynamicCache. Because BioGPT uses regular multi-head attention, KV heads equal attention heads and head_dim is 1024 / 16 = 64." }, "notes": "BioGPT is a dense text-only decoder model. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "fp32", "kv_store_bytes_per_scalar": 4, "kv_read_format": "fp32", "kv_read_bytes_per_scalar": 4, "runtime_format": "transformers-fp32-biogpt-text-decode-memory-bound", "dequantization_notes": "No quantized or reduced-precision serving representation is assumed for the selected PyTorch checkpoint. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The config does not declare torch_dtype, and the pinned PyTorch checkpoint metadata stores all tensors as FloatStorage. KV cache is charged at FP32 four bytes per scalar to match the default loaded parameter dtype." }, "evidence": [ { "label": "BioGPT API metadata", "url": "https://huggingface.co/api/models/microsoft/biogpt", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "serving" ], "notes": "At repo SHA eb0d815e95434dc9e3b78f464e52b899bee7d923, the API records a public/non-gated MIT Transformers/PyTorch text-generation repo with biogpt, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 122777. The API does not expose safetensors or GGUF tensor metadata because this is a legacy PyTorch checkpoint repo." }, { "label": "BioGPT model card", "url": "https://huggingface.co/microsoft/biogpt/raw/eb0d815e95434dc9e3b78f464e52b899bee7d923/README.md", "source_type": "model_card", "supports": [ "license", "pipeline", "model_family" ], "notes": "The model card records MIT licensing, English language metadata, text-generation usage through BioGptForCausalLM.from_pretrained(\"microsoft/biogpt\"), and examples using max_length up to 1024." }, { "label": "BioGPT config", "url": "https://huggingface.co/microsoft/biogpt/raw/eb0d815e95434dc9e3b78f464e52b899bee7d923/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "head_dim", "max_context_tokens", "kv_adapter" ], "notes": "The config records BioGptForCausalLM, model_type biogpt, hidden_size 1024, intermediate_size 4096, num_hidden_layers 24, num_attention_heads 16, max_position_embeddings 1024, vocab_size 42384, use_cache true, and scale_embedding true." }, { "label": "Transformers BioGPT implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/08a7ef05bcf9723cb2e58855afb8dc2c799323ff/src/transformers/models/biogpt/modeling_biogpt.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "embedding_layout", "position_embedding_layout" ], "notes": "Manual review found BioGptAttention builds q_proj, k_proj, v_proj, and out_proj, reshapes K/V by head_dim, and writes key_states and value_states into DynamicCache. BioGptModel constructs embed_tokens, BioGptLearnedPositionalEmbedding with a two-row offset, decoder layers, and a final layer_norm. BioGptForCausalLM declares output_projection.weight tied to biogpt.embed_tokens.weight." }, { "label": "BioGPT PyTorch checkpoint metadata audit", "url": "https://huggingface.co/microsoft/biogpt/resolve/eb0d815e95434dc9e3b78f464e52b899bee7d923/pytorch_model.bin", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving" ], "notes": "HEAD checks resolve pytorch_model.bin to a 1560781537-byte linked object. Range-reading the PyTorch zip data.pkl found 389 serialized FloatStorage tensors. Serialized tensor params total 390164480 FP32 scalars: biogpt.embed_tokens.weight 43401216, biogpt.embed_positions.weight 1050624, 24 decoder layers totaling 302309376, biogpt.layer_norm.* 2048, and duplicate output_projection.weight 43401216. Loading with Transformers ties output_projection.weight to biogpt.embed_tokens.weight, so unique serving resident params are 346763264 FP32 scalars, ordinary text swept params are 345712640, and resident-only position params are 1050624." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned config, local Transformers BioGPT implementation at commit 08a7ef05bcf9723cb2e58855afb8dc2c799323ff, PyTorch linked-object HEAD metadata, and range-read PyTorch zip data.pkl tensor metadata." }, "notes": "This profile supersedes the scraped metadata estimate by using the selected PyTorch checkpoint's FP32 dtype, exact tied serving parameter count, exact resident-only learned position embedding bytes, and full-context BioGPT KV geometry." }, { "id": "microsoft--dialogpt-medium", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/DialoGPT-medium", "title": "Microsoft DialoGPT Medium FP16", "summary": "Audited memory-side text-decode bounds profile for the FP16 PyTorch DialoGPT-medium repo.", "model_family": "gpt2-dense", "base_model_proof": { "base_model": "microsoft/DialoGPT-medium", "relation": "base", "source": "Hugging Face model metadata, pinned GPT-2 config, model card, and direct PyTorch checkpoint load", "config_compatible": true, "notes": "DialoGPT-medium is the target repo itself. The config records GPT2LMHeadModel with GPT-2 medium geometry, and the pinned PyTorch checkpoint loads into AutoModelForCausalLM without needing a derived architecture." }, "architecture": { "canonical_architecture_id": "gpt2-medium", "max_context_tokens": 1024, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.354823168, "swept_params_b": 0.353774592, "auxiliary_resident_params_b": 0.001048576, "resident_weight_gb": 0.709646336, "swept_weight_gb": 0.707549184, "auxiliary_resident_weight_gb": 0.002097152, "resident_parameter_scope": "unique FP16 parameters after Transformers GPT-2 weight tying", "swept_parameter_scope": "ordinary text decode sweeps transformer.h.*, transformer.ln_f.*, and the tied transformer.wte/lm_head output matrix", "auxiliary_scope": "transformer.wpe.weight is resident for position lookup but is not swept as a full matrix for each ordinary generated token", "notes": "The raw PyTorch state dict stores 317 FP16 tensors totaling 0.862904320 GB, including duplicate lm_head.weight and persistent transformer.h.*.attn.bias causal-mask buffers. Loading the pinned checkpoint with Transformers 4.57.1 ties lm_head.weight to transformer.wte.weight and leaves no named buffers in the serving model, producing 292 unique FP16 parameters totaling 0.709646336 GB. This profile charges serving-resident unique parameter bytes, not serialized checkpoint duplicates." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 24 layers, 16 attention heads, hidden size 1024, and GPT-2 multi-head attention, so key and value caches each use 16 heads * 64 dimensions per layer." }, "notes": "DialoGPT-medium is a dense text-only GPT-2 decoder. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp16-gpt2-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected PyTorch checkpoint. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The pinned PyTorch checkpoint loads all serving parameters as torch.float16. KV cache is charged at FP16 two bytes per scalar to match the loaded parameter dtype." }, "evidence": [ { "label": "DialoGPT-medium API metadata", "url": "https://huggingface.co/api/models/microsoft/DialoGPT-medium", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "serving" ], "notes": "At repo SHA 7b40bb0f92c45fefa957d088000d8648e5c7fa33, the API records a public/non-gated MIT text-generation repo with GPT-2, conversational, PyTorch, TensorFlow, JAX, Rust, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 233570. The API does not expose safetensors metadata because this is a legacy checkpoint repo." }, { "label": "DialoGPT-medium model card", "url": "https://huggingface.co/microsoft/DialoGPT-medium/raw/7b40bb0f92c45fefa957d088000d8648e5c7fa33/README.md", "source_type": "model_card", "supports": [ "license", "pipeline", "model_family" ], "notes": "The model card records MIT licensing, conversational tagging, DialoGPT as a large-scale pretrained response generation model, and usage through AutoModelForCausalLM.from_pretrained(\"microsoft/DialoGPT-medium\")." }, { "label": "DialoGPT-medium config", "url": "https://huggingface.co/microsoft/DialoGPT-medium/raw/7b40bb0f92c45fefa957d088000d8648e5c7fa33/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "head_dim", "max_context_tokens", "kv_adapter" ], "notes": "The config records GPT2LMHeadModel, model_type gpt2, n_ctx and n_positions 1024, n_embd 1024, n_head 16, n_layer 24, vocab_size 50257, activation gelu_new, and conversational task max_length 1000." }, { "label": "Transformers 4.57.1 GPT-2 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.1/src/transformers/models/gpt2/modeling_gpt2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "embedding_layout" ], "notes": "Manual review found GPT2Attention splits c_attn output into query_states, key_states, and value_states, reshapes K/V by self.head_dim, and writes key_states and value_states into the cache. GPT2Model instantiates wte and wpe embeddings. GPT2LMHeadModel declares _tied_weights_keys [\"lm_head.weight\"], constructs lm_head, and computes logits through lm_head." }, { "label": "DialoGPT-medium PyTorch checkpoint audit", "url": "https://huggingface.co/microsoft/DialoGPT-medium/resolve/7b40bb0f92c45fefa957d088000d8648e5c7fa33/pytorch_model.bin", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving" ], "notes": "HEAD checks resolve pytorch_model.bin to 862955157 bytes. Direct torch.load of the pinned file found 317 serialized FP16 tensors totaling 0.862904320 GB: transformer.wte.weight 0.102926336 GB, transformer.wpe.weight 0.002097152 GB, transformer.h.* non-mask tensors 0.604618752 GB, transformer.h.*.attn.bias causal-mask buffers 0.050331648 GB, transformer.ln_f.* 0.000004096 GB, and lm_head.weight 0.102926336 GB. Loading the same pinned directory with Transformers 4.57.1 AutoModelForCausalLM ties lm_head.weight to transformer.wte.weight and leaves zero named buffers, yielding 292 unique FP16 serving parameters totaling 0.709646336 GB. The swept ordinary text subset is transformer.h.* non-mask tensors plus transformer.ln_f.* plus the tied wte/lm_head output matrix: 0.707549184 GB. The only resident-only serving parameter is transformer.wpe.weight: 0.002097152 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, immutable config, Transformers 4.57.1 GPT-2 implementation, pinned PyTorch checkpoint tensor grouping, and actual AutoModelForCausalLM load-time weight tying behavior." }, "notes": "This profile supersedes the scraped metadata estimate by using the selected PyTorch checkpoint's FP16 dtype and by separating the tied output embedding sweep from resident-only position embedding bytes." }, { "id": "microsoft--florence-2-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "microsoft/Florence-2-base", "title": "Microsoft Florence 2 Base FP16", "summary": "Unsupported profile stub for the FP16 Florence-2 Base encoder-decoder vision-language repo.", "model_family": "florence2-encoder-decoder", "architecture": { "canonical_architecture_id": "florence-2-base", "max_context_tokens": 1024, "weight_adapter": { "kind": "dense", "total_params_b": 0.231567705, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API and direct safetensors header both record 231567705 FP16 parameters. Production bounds are disabled because the architecture is encoder-decoder with vision inputs, not an ordinary decoder-only text model." }, "kv_adapter": { "kind": "unknown", "reason": "The config records Florence2ForConditionalGeneration with is_encoder_decoder true, a DaViT vision tower, a text encoder, a text decoder, and decoder encoder_attn cross-attention weights. Bounds Engine v1 has no adapter for encoder-decoder cross-attention state, encoder output traffic, or vision encoder throughput.", "notes": "Do not infer an ordinary full-context decoder KV adapter from decoder_layers and decoder_attention_heads alone. A future Florence adapter must separately account for decoder self-attention KV, cross-attention over encoded image/text state, and vision encoder work." }, "notes": "The audited config exposes enough geometry to reject this model for Bounds Engine v1 without guessing." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-encoder-decoder-vision-language", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because the runtime state traffic is outside the supported decoder-only adapter set.", "notes": "FP16 weight dtype comes from the Hugging Face API safetensors metadata and direct safetensors header." }, "evidence": [ { "label": "Microsoft Florence-2 Base API metadata", "url": "https://huggingface.co/api/models/microsoft/Florence-2-base", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "license", "pipeline" ], "notes": "The API reports a public MIT-licensed image-text-to-text repo with custom Florence2 code and FP16 safetensors total 231567705." }, { "label": "Microsoft Florence-2 Base config", "url": "https://huggingface.co/microsoft/Florence-2-base/raw/main/config.json", "source_type": "config", "supports": [ "architecture", "max_context_tokens", "unsupported_reason" ], "notes": "The config records architectures Florence2ForConditionalGeneration, is_encoder_decoder true, torch_dtype float16, text_config max_position_embeddings 1024, six encoder layers, six decoder layers, twelve attention heads, and a DaViT vision_config." }, { "label": "Microsoft Florence-2 Base safetensors header audit", "url": "https://huggingface.co/microsoft/Florence-2-base/resolve/main/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "serving", "unsupported_reason" ], "notes": "Header-only audit found one 463221266-byte safetensors file, 85848-byte header, 666 tensors, 463135410 FP16 tensor bytes, language_model tensor bytes 280516274, vision_tower tensor bytes 180736000, and decoder encoder_attn tensors." } ], "unsupported_reason": "Florence-2 Base is an encoder-decoder vision-language model with decoder cross-attention over encoded state. Bounds Engine v1 only supports audited decoder-side memory traffic adapters, so production tok/s bounds are disabled until an explicit encoder-decoder adapter exists.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports Florence-style encoder-decoder and vision state traffic." }, { "id": "microsoft--florence-2-large", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "microsoft/Florence-2-large", "title": "Microsoft Florence 2 Large FP16", "summary": "Unsupported profile stub for the FP16 Florence-2 Large encoder-decoder vision-language repo.", "model_family": "florence2-encoder-decoder", "architecture": { "canonical_architecture_id": "florence-2-large", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense", "total_params_b": 0.776721497, "parameter_scope": "hf_api_and_safetensors_header_total", "notes": "The Hugging Face API and direct safetensors header both record 776721497 FP16 parameters. Production bounds are disabled because the architecture is encoder-decoder with vision inputs, not an ordinary decoder-only text model." }, "kv_adapter": { "kind": "unknown", "reason": "The config records Florence2ForConditionalGeneration with is_encoder_decoder true, a DaViT vision tower, a text encoder, a text decoder, and decoder encoder_attn cross-attention weights. Bounds Engine v1 has no adapter for encoder-decoder cross-attention state, encoder output traffic, or vision encoder throughput.", "notes": "Do not infer an ordinary full-context decoder KV adapter from decoder_layers and decoder_attention_heads alone. A future Florence adapter must separately account for decoder self-attention KV, cross-attention over encoded image/text state, and vision encoder work." }, "notes": "The audited config exposes enough geometry to reject this model for Bounds Engine v1 without guessing." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-encoder-decoder-vision-language", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because the runtime state traffic is outside the supported decoder-only adapter set.", "notes": "FP16 weight dtype comes from the Hugging Face API safetensors metadata and direct safetensors header." }, "evidence": [ { "label": "Microsoft Florence-2 Large API metadata", "url": "https://huggingface.co/api/models/microsoft/Florence-2-large", "source_type": "model_card", "supports": [ "repo", "total_params_b", "serving", "license", "pipeline" ], "notes": "At commit 21a599d414c4d928c9032694c424fb94458e3594, the API reports a public MIT-licensed image-text-to-text repo with Transformers custom Florence2 code, region:us, current downloads 662200, and FP16 safetensors total 776721497." }, { "label": "Microsoft Florence-2 Large config", "url": "https://huggingface.co/microsoft/Florence-2-large/raw/21a599d414c4d928c9032694c424fb94458e3594/config.json", "source_type": "config", "supports": [ "architecture", "max_context_tokens", "unsupported_reason" ], "notes": "The config records architectures Florence2ForConditionalGeneration, is_encoder_decoder true, torch_dtype float16, text_config max_position_embeddings 4096, d_model 1024, twelve encoder layers, twelve decoder layers, sixteen encoder and decoder attention heads, 4096 FFN width, 51289 vocab size, and a DaViT vision_config with depths [1, 1, 9, 1], dim_embed [256, 512, 1024, 2048], and projection_dim 1024." }, { "label": "Microsoft Florence-2 Large safetensors header audit", "url": "https://huggingface.co/microsoft/Florence-2-large/resolve/21a599d414c4d928c9032694c424fb94458e3594/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "serving", "unsupported_reason" ], "notes": "Header-only audit found one safetensors file with a 120456-byte header and 918 FP16 tensors totaling 776721497 parameters / 1.553442994 GB. Grouped payloads: vision_tower 360632320 params / 0.72126464 GB, language encoder 155353088 params / 0.310706176 GB, language decoder 205758464 params / 0.411516928 GB, shared embedding 52519936 params / 0.105039872 GB, language_other 51289 params / 0.000102578 GB, and other tensors 2406400 params / 0.0048128 GB. Decoder encoder_attn cross-attention tensors contribute 50380800 params / 0.1007616 GB." } ], "unsupported_reason": "Florence-2 Large is an encoder-decoder vision-language model with decoder cross-attention over encoded state. Bounds Engine v1 only supports audited decoder-side memory traffic adapters, so production tok/s bounds are disabled until an explicit encoder-decoder adapter exists.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports Florence-style encoder-decoder and vision state traffic." }, { "id": "microsoft--kosmos-2-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "microsoft/kosmos-2.5", "title": "Microsoft Kosmos 2.5 F32", "summary": "Unsupported profile stub with exact resident tensor evidence for the F32 Kosmos 2.5 image-to-text document model.", "model_family": "kosmos-2.5-multimodal-document", "architecture": { "canonical_architecture_id": "kosmos-2.5", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense", "total_params_b": 1.374646272, "parameter_scope": "hf_api_safetensors_total_f32", "notes": "The Hugging Face API records 1374646272 F32 safetensors parameters. Direct shard-header range reads found 5.498585088 GB of F32 tensor payload split into 3.387537408 GB text_model tensors, 2.051248128 GB vision_model tensors, and 0.059799552 GB image_to_text_projection tensors." }, "kv_adapter": { "kind": "unknown", "reason": "Kosmos 2.5 is an image-text-to-text document/OCR model with flattened image patches, a vision tower, image-to-text projection tensors, and 2048 latent queries. Bounds Engine v1 does not model image-conditioned OCR/markdown generation state, visual patch prefill, latent query traffic, or decoder interaction with projected image features.", "notes": "The config exposes a 24-layer text_config with use_cache true, but using that alone as an ordinary text-only KV adapter would omit the image-conditioned state that defines the production workload." }, "notes": "This profile intentionally fails closed even though tensor bytes and text/vision grouping are accessible. The current bounds engine needs an explicit multimodal document/OCR adapter before this repo can show production tok/s estimates." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 4, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 4, "runtime_format": "unsupported-kosmos2-5-image-conditioned-document-generation", "dequantization_notes": "No quantized weight representation is assumed for the stored F32 safetensors. The model card examples load the model with dtype bfloat16 for inference, but the repository tensors are stored as F32.", "notes": "Production OCR/markdown throughput depends on image patch count, visual encoder execution, projected image features, task prompt, and output length. Bounds Engine v1 only supports audited text-token decode adapters." }, "evidence": [ { "label": "Kosmos 2.5 API metadata", "url": "https://huggingface.co/api/models/microsoft/kosmos-2.5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "serving", "total_params_b", "unsupported_reason" ], "notes": "At commit ec3c8051b697166514a31d646cfa36d6ef4c93d7, the current API records a public non-gated MIT image-text-to-text repo with transformers, safetensors, kosmos-2.5, en, arXiv 2309.11419, endpoints_compatible, and region:us tags. Current downloads are 100827. The API safetensors block records F32 1374646272 and total 1374646272." }, { "label": "Kosmos 2.5 model card", "url": "https://huggingface.co/microsoft/kosmos-2.5/raw/ec3c8051b697166514a31d646cfa36d6ef4c93d7/README.md", "source_type": "model_card", "supports": [ "repo", "pipeline", "runtime_format", "unsupported_reason" ], "notes": "The model card describes Kosmos 2.5 as a multimodal literate model for machine reading of text-intensive images, with OCR and image-to-markdown tasks. The examples run Kosmos2_5ForConditionalGeneration with an AutoProcessor, image inputs, flattened_patches converted to bfloat16, and task prompts such as and ." }, { "label": "Kosmos 2.5 config", "url": "https://huggingface.co/microsoft/kosmos-2.5/raw/ec3c8051b697166514a31d646cfa36d6ef4c93d7/config.json", "source_type": "config", "supports": [ "model_family", "text_config", "vision_config", "max_context_tokens", "unsupported_reason" ], "notes": "The config records Kosmos2_5ForConditionalGeneration, model_type kosmos-2.5, top-level torch_dtype float32, latent_query_num 2048, a text_config with 24 layers, 1536 embed_dim, 16 attention heads, 6144 FFN dimension, max_position_embeddings 4096, use_cache true, and tied text embeddings; and a vision_config with 18 layers, hidden size 1536, 24 attention heads, head_dim 64, patch_embed_hidden_size 768, max_num_patches 4096, and max_length 4096." }, { "label": "Kosmos 2.5 safetensors index and shard headers", "url": "https://huggingface.co/microsoft/kosmos-2.5/raw/ec3c8051b697166514a31d646cfa36d6ef4c93d7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "weight_format", "text_vision_split", "unsupported_reason" ], "notes": "The index records total_size 5498585088 bytes across two safetensors shards. Direct range-read shard headers found 614 F32 tensors totaling 5.498585088 GB, with linked file size 5.498660528 GB including 0.000075440 GB of safetensors header/container bytes. Tensor groups are text_model 3.387537408 GB, vision_model 2.051248128 GB, and image_to_text_projection 0.059799552 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from current HF API metadata, pinned model card, served config, safetensors index, and direct safetensors shard-header range reads." }, "unsupported_reason": "Kosmos 2.5 is an image-conditioned document OCR/markdown generation model. Bounds Engine v1 lacks an adapter for visual patch prefill, latent queries, projected image features, and their interaction with generated text, so ordinary text-only tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports a Kosmos-style image-conditioned document generation workload adapter." }, { "id": "microsoft--phi-2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/phi-2", "title": "Microsoft Phi-2 F16", "summary": "Audited memory-side text-decode bounds profile for the F16 Phi-2 repo.", "model_family": "phi-dense", "architecture": { "canonical_architecture_id": "phi-2", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.77968384, "swept_params_b": 2.64861184, "auxiliary_resident_params_b": 0.131072, "resident_weight_gb": 5.55936768, "swept_weight_gb": 5.29722368, "auxiliary_resident_weight_gb": 0.262144, "resident_parameter_scope": "safetensors_header_stored_f16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup while including model.layers.*, model.final_layernorm.*, lm_head.weight, and lm_head.bias", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain separate lm_head.weight and lm_head.bias tensors, so the output head is swept while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are F16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 80, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 32 attention heads, 32 KV heads, hidden size 2560, head_dim 80, and 2048 max position embeddings. The official Transformers 4.37.0 Phi implementation builds separate q_proj, k_proj, and v_proj layers, then caches key_states and value_states." }, "notes": "PhiForCausalLM is a dense text-only decoder model. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp16-phi-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this F16 repo. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype float16, and range-read safetensors shard headers record only F16 tensors. KV cache is charged at FP16 two bytes per scalar." }, "evidence": [ { "label": "Phi-2 model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/phi-2", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA 810d367871c1d460086d9f82db8696f2e0a0fcd0, the API records a public/non-gated MIT text-generation repo with phi, safetensors, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 787841. The API safetensors block reports F16: 2779683840 and total: 2779683840. The model card describes Phi-2 as a 2.7B Transformer with 2048-token context and MIT licensing." }, { "label": "Phi-2 config", "url": "https://huggingface.co/microsoft/phi-2/raw/810d367871c1d460086d9f82db8696f2e0a0fcd0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records PhiForCausalLM, model_type phi, torch_dtype float16, tie_word_embeddings false, 32 decoder layers, hidden size 2560, intermediate size 10240, 32 attention heads, 32 KV heads, 2048 max position embeddings, vocab size 51200, partial_rotary_factor 0.4, rope_theta 10000, and qk_layernorm false." }, { "label": "Transformers 4.37.0 Phi implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.37.0/src/transformers/models/phi/modeling_phi.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "embedding_layout" ], "notes": "Manual review found PhiAttention instantiates q_proj, k_proj, and v_proj separately with k/v output size num_key_value_heads * head_dim, then writes key_states and value_states into past_key_value.update. PhiModel instantiates model.embed_tokens, and PhiForCausalLM instantiates a separate lm_head linear layer used for logits." }, { "label": "Phi-2 safetensors index and shard headers", "url": "https://huggingface.co/microsoft/phi-2/resolve/810d367871c1d460086d9f82db8696f2e0a0fcd0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 5559367680 bytes across two shards. Range-read shard headers found 453 F16 tensors totaling 2.779683840B parameters / 5.559367680 GB. model.layers.* tensors total 2.517483520B parameters / 5.034967040 GB, model.final_layernorm.* is 5120 parameters / 0.000010240 GB, lm_head.weight plus lm_head.bias is 131.123200M parameters / 0.262246400 GB, and model.embed_tokens.weight is 131.072000M parameters / 0.262144000 GB. The swept ordinary text subset is model.layers.* plus model.final_layernorm.* plus lm_head.*: 2.648611840B parameters / 5.297223680 GB. The resident-only subset is model.embed_tokens.weight: 0.131072000B parameters / 0.262144000 GB." }, { "label": "Phi-2 license", "url": "https://huggingface.co/microsoft/phi-2/raw/810d367871c1d460086d9f82db8696f2e0a0fcd0/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, official Transformers 4.37.0 Phi implementation, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights." }, { "id": "microsoft--phi-3-5-mini-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3.5-mini-instruct", "title": "Microsoft Phi-3.5 Mini Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-3.5 Mini Instruct repo.", "model_family": "phi3-dense", "architecture": { "canonical_architecture_id": "phi-3-5-mini", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.821079552, "swept_params_b": 3.722578944, "auxiliary_resident_params_b": 0.098500608, "resident_weight_gb": 7.642159104, "swept_weight_gb": 7.445157888, "auxiliary_resident_weight_gb": 0.197001216, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 96, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 32 attention heads, 32 KV heads, hidden size 3072, and max context 131072. The remote code builds qkv_proj with hidden_size to num_heads*head_dim plus two KV streams. Config sliding_window is 262144, above max_position_embeddings, so it does not cap the profile's supported context." }, "notes": "Phi3ForCausalLM is a dense text-only decoder model. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-3.5 Mini Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-mini-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA 2fe192450127e6a83f7441aef6e3ca586c338b77, the API records a public/non-gated MIT text-generation repo with phi3, custom_code, multilingual, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 1030083. The API safetensors block reports BF16: 3821079552 and total: 3821079552. The card describes Phi-3.5 Mini as a dense 3.8B decoder-only Transformer with 128K context support and MIT licensing." }, { "label": "Phi-3.5 Mini Instruct config", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 131072 max position embeddings, original 4096 context, LongRoPE scaling, sliding_window 262144, and vocab size 32064." }, { "label": "Phi-3.5 Mini Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/modeling_phi3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "sliding_window_boundary" ], "notes": "Manual review found Phi3Attention builds qkv_proj with output size num_heads*head_dim plus two num_key_value_heads*head_dim streams, then caches key_states and value_states. Phi3ForCausalLM instantiates a separate lm_head linear layer. FlashAttention sliding-window support only applies when kv_seq_len exceeds config.sliding_window, which is above this repo's max_position_embeddings." }, { "label": "Phi-3.5 Mini Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/resolve/2fe192450127e6a83f7441aef6e3ca586c338b77/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 7642159104 bytes across two shards. Range-read shard headers found 195 BF16 tensors totaling 3821079552 parameters / 7.642159104 GB. model.layers.* tensors total 3.624075264B parameters / 7.248150528 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, lm_head.weight is 98.500608M parameters / 0.197001216 GB, and model.embed_tokens.weight is 98.500608M parameters / 0.197001216 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 3.722578944B parameters / 7.445157888 GB. The resident-only subset is model.embed_tokens.weight: 0.098500608B parameters / 0.197001216 GB." }, { "label": "Phi-3.5 Mini Instruct license", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights." }, { "id": "microsoft--phi-3-5-moe-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3.5-MoE-instruct", "title": "Microsoft Phi-3.5 MoE Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for Microsoft's BF16 Phi-3.5-MoE-Instruct repo.", "model_family": "phi-3.5-moe", "architecture": { "canonical_architecture_id": "phi-3.5-moe", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 83.746306688, "main_resident_weight_gb": 83.4836384, "auxiliary_resident_weight_gb": 0.262668288, "fixed_weight_gb": 2.9530016, "routed_expert_weight_gb": 5.0331648, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all 32 decoder layers, model.norm.weight, lm_head.weight, and lm_head.bias, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records no shared expert field; every decoder layer uses 16 routed local experts and top-2 sparse routing.", "notes": "Header-derived BF16 bytes are used instead of rounded model-card counts. Routed expert tensors are byte-uniform across all 16 expert indexes; routed_expert_weight_gb is the total expert tensor byte count divided by 16. Fixed ordinary-decode traffic includes attention projections, routers, layer norms, final norm, and the untied LM head." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 decoder layers, 8 KV heads, 128-dimensional key/value heads, 131072 max positions, and sliding_window 131072. Because the sliding window equals the advertised maximum context, the ordinary max-context serving profile is equivalent to full-context KV. The custom attention code caches key_states and value_states through DynamicCache when use_cache is enabled." }, "notes": "PhiMoEForCausalLM decoder-only sparse MoE profile with 16 local experts per layer and top-2 routing." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phimoe-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert kernel behavior, FlashAttention efficiency, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16 and the model card's local loading example uses torch_dtype auto. KV cache is charged at BF16 two bytes per scalar for ordinary cached serving." }, "evidence": [ { "label": "Phi-3.5-MoE-Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-MoE-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "At repo SHA 43688451b462a3351d8580625ebe1931adb3986d, the API records a public/non-gated MIT Transformers text-generation repo with phimoe, custom_code, safetensors, and region:us tags. Current downloads are 162753. The API safetensors block reports BF16: 41873153344 and total: 41873153344. The model card describes 128K context, 16x3.8B parameters, and 6.6B active parameters with 2 experts." }, { "label": "Phi-3.5-MoE-Instruct config", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/raw/43688451b462a3351d8580625ebe1931adb3986d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records PhiMoEForCausalLM, model_type phimoe, BF16 dtype, 32 decoder layers, hidden size 4096, intermediate size 6400, 32 attention heads, 8 KV heads, 131072 max position embeddings, LongRoPE from 4096 original positions, sliding_window 131072, 16 local experts, 2 experts per token, attention/lm_head bias true, tie_word_embeddings false, use_cache true, and vocab size 32064." }, { "label": "Phi-3.5-MoE-Instruct custom modeling file", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/raw/43688451b462a3351d8580625ebe1931adb3986d/modeling_phimoe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "ordinary_decode_scope" ], "notes": "Manual review found PhiMoEAttention projecting and caching key_states and value_states with num_key_value_heads x head_dim through DynamicCache, PhiMoESparseMoeBlock building 16 local experts and selecting top-2 experts, PhiMoEModel using model.embed_tokens and 32 PhiMoEDecoderLayer blocks, and PhiMoEForCausalLM using a separate biased lm_head." }, { "label": "Phi-3.5-MoE-Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/resolve/43688451b462a3351d8580625ebe1931adb3986d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 83746306688 bytes across 17 shards. Range-read safetensors headers found 1957 BF16 tensors totaling 83.746306688 GB. model.embed_tokens.weight is 0.262668288 GB resident-only. Ordinary main tensors excluding input embeddings sum to 83.483638400 GB. Routed expert tensors in model.layers.*.block_sparse_moe.experts.* sum to 80.530636800 GB, exactly 5.033164800 GB per expert index. Fixed ordinary-decode traffic, including attention projections, routers, layer norms, model.norm.weight, lm_head.weight, and lm_head.bias, is 2.953001600 GB." }, { "label": "Phi-3.5-MoE-Instruct license", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/raw/43688451b462a3351d8580625ebe1931adb3986d/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned config, pinned custom modeling code, safetensors index, direct range-read safetensors shard headers, and license file." }, "notes": "This profile supersedes the generated metadata estimate, which rounded total storage and undercounted active MoE traffic by omitting fixed per-token non-expert traffic." }, { "id": "microsoft--phi-3-5-vision-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3.5-vision-instruct", "title": "Microsoft Phi-3.5 Vision Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-3.5 Vision Instruct repo.", "model_family": "phi3-v-dense", "architecture": { "canonical_architecture_id": "phi-3-5-vision", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.14662144, "swept_params_b": 3.722578944, "auxiliary_resident_params_b": 0.424042496, "resident_weight_gb": 8.29324288, "swept_weight_gb": 7.445157888, "auxiliary_resident_weight_gb": 0.848084992, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and resident vision embedding/CLIP tensors while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.vision_embed_tokens.* and model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 96, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 32 attention heads, 32 KV heads, hidden size 3072, and max context 131072. The remote code builds qkv_proj with hidden_size to num_heads*head_dim plus two KV streams. Config sliding_window is 262144, above max_position_embeddings, so it does not cap the profile's supported context." }, "notes": "Phi3VForCausalLM is multimodal with a resident CLIP vision embedding path. This profile models ordinary text decode after any image/video prefill, not CLIP vision encoder or image projection throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3v-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, image projection, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-3.5 Vision Instruct model card", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "max_context_tokens", "multimodal" ], "notes": "The card records MIT licensing, image-text-to-text packaging, custom Transformers code, and describes Phi-3.5 Vision as a lightweight multimodal model with 128K context support." }, { "label": "Phi-3.5 Vision Instruct HF CLI/API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-vision-instruct", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "weight_format", "total_params_b", "safetensors_total_size" ], "notes": "The live HF CLI/API response at commit 12b77fb40b63a2c73c68243d3f767aab688a1b2a records downloads 1300939, safetensors parameters BF16: 4146621440, total: 4146621440, and storage 8293330888 bytes including safetensors headers." }, { "label": "Phi-3.5 Vision Instruct config", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Phi3VForCausalLM, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 131072 max position embeddings, original 4096 context, SU rope scaling, sliding_window 262144, vocab size 32064, CLIP vision model openai/clip-vit-large-patch14-336, image_dim_out 1024, and image MLP projection settings." }, { "label": "Phi-3.5 Vision Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/modeling_phi3_v.py", "source_type": "manual_review", "supports": [ "kv_adapter", "vision_embedding_scope", "lm_head_layout" ], "notes": "Manual review found Phi3Attention builds qkv_proj with output size num_heads*head_dim + 2*num_key_value_heads*head_dim, Phi3VModel instantiates model.vision_embed_tokens from Phi3ImageEmbedding when embd_layer is present, and Phi3VForCausalLM instantiates a separate lm_head linear layer." }, { "label": "Phi-3.5 Vision Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_resident_scope" ], "notes": "The index records metadata total_size 8293242880 bytes across two shards. HEAD checks report linked shard sizes 4944122112 and 3349208776 bytes; the 88008-byte difference is safetensors header overhead, matching HF used_storage. Range-read shard headers found 592 BF16 tensors totaling 4146621440 parameters / 8.29324288 GB. model.layers.* tensors total 3.624075264B parameters / 7.248150528 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, lm_head.weight is 98.500608M parameters / 0.197001216 GB, model.embed_tokens.weight is 98.500608M parameters / 0.197001216 GB, and model.vision_embed_tokens.* tensors total 325.541888M parameters / 0.651083776 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 3.722578944B parameters / 7.445157888 GB. The resident-only subset is model.embed_tokens.weight plus model.vision_embed_tokens.*: 0.424042496B parameters / 0.848084992 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF CLI/API metadata, model card, immutable config, remote modeling code, safetensors index, HEAD checks, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only vision and input-embedding tensors from per-token swept text/logit weights." }, { "id": "microsoft--phi-3-mini-128k-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3-mini-128k-instruct", "title": "Microsoft Phi-3 Mini 128K Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-3 Mini 128K Instruct repo.", "model_family": "phi3-dense", "architecture": { "canonical_architecture_id": "phi-3-mini-128k", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.821079552, "swept_params_b": 3.722578944, "auxiliary_resident_params_b": 0.098500608, "resident_weight_gb": 7.642159104, "swept_weight_gb": 7.445157888, "auxiliary_resident_weight_gb": 0.197001216, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 96, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 32 attention heads, 32 KV heads, hidden size 3072, 131072 max position embeddings, and sliding_window 262144. Because the sliding window is above max_position_embeddings, it does not cap this profile's supported context." }, "notes": "Phi3ForCausalLM is a dense text-only decoder model. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, LongRoPE math, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-3 Mini 128K Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-mini-128k-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA f3c06aed622e14ca0abf5115094e4fc9a9948f36, the API records a public/non-gated MIT text-generation repo with phi3, custom_code, code, conversational, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 243935. The API safetensors block reports BF16: 3821079552 and total: 3821079552. The model card describes Phi-3 Mini 128K Instruct as a dense 3.8B decoder-only model with 128K context support." }, { "label": "Phi-3 Mini 128K Instruct config", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/raw/f3c06aed622e14ca0abf5115094e4fc9a9948f36/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 131072 max position embeddings, original 4096 context, LongRoPE scaling, sliding_window 262144, and vocab size 32064." }, { "label": "Phi-3 Mini 128K Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/raw/f3c06aed622e14ca0abf5115094e4fc9a9948f36/modeling_phi3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "sliding_window_boundary" ], "notes": "Manual review found Phi3Attention builds qkv_proj with output size num_heads*head_dim plus two num_key_value_heads*head_dim streams, then caches key_states and value_states. Phi3ForCausalLM instantiates a separate lm_head linear layer. FlashAttention sliding-window support only applies when kv_seq_len exceeds config.sliding_window, which is above this repo's max_position_embeddings." }, { "label": "Phi-3 Mini 128K Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/f3c06aed622e14ca0abf5115094e4fc9a9948f36/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 7642159104 bytes across two shards. Range-read shard headers found 195 BF16 tensors totaling 3821079552 parameters / 7.642159104 GB. model.layers.* tensors total 3.624075264B parameters / 7.248150528 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, lm_head.weight is 98.500608M parameters / 0.197001216 GB, and model.embed_tokens.weight is 98.500608M parameters / 0.197001216 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 3.722578944B parameters / 7.445157888 GB. The resident-only subset is model.embed_tokens.weight: 0.098500608B parameters / 0.197001216 GB. Linked-object HEAD checks resolved both shard files to 7.642181880 GB total, leaving 0.000022776 GB of safetensors header/container overhead outside tensor payloads." }, { "label": "Phi-3 Mini 128K Instruct license", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/raw/f3c06aed622e14ca0abf5115094e4fc9a9948f36/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." }, { "label": "Phi-3.5 Mini Instruct profile comparison", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/config.json", "source_type": "manual_review", "supports": [ "model_family", "weight_adapter", "kv_adapter" ], "notes": "Manual comparison against the already audited Phi-3.5 Mini Instruct profile found matching text geometry, tensor names, stored BF16 tensor byte totals, separate input embedding and lm_head layout, full-context KV behavior, and 131072 max positions. The LongRoPE factor arrays differ, but those do not change Bounds Engine v1 memory traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, direct safetensors shard header byte grouping, and comparison against the already audited Phi-3.5 Mini Instruct profile." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights and by charging the full 128K BF16 K/V cache rather than applying the 4K sibling's sliding-window cap." }, { "id": "microsoft--phi-3-mini-4k-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3-mini-4k-instruct", "title": "Microsoft Phi-3 Mini 4K Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-3 Mini 4K Instruct repo.", "model_family": "phi3-dense", "architecture": { "canonical_architecture_id": "phi-3-mini-4k", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.821079552, "swept_params_b": 3.722578944, "auxiliary_resident_params_b": 0.098500608, "resident_weight_gb": 7.642159104, "swept_weight_gb": 7.445157888, "auxiliary_resident_weight_gb": 0.197001216, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 32, "kv_heads": 32, "head_dim": 96, "window_tokens": 2047, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window 2047. The remote FlashAttention path slices cached K/V tensors to sliding_window - 1 before appending the next token, so the profile caps allocation and read traffic at 2047 live K/V tokens per layer." } ], "notes": "All 32 layers use the same sliding-window K/V geometry." }, "notes": "Phi3ForCausalLM is a dense text-only decoder model. This profile models ordinary autoregressive text decode with the repo's FlashAttention sliding-window cache behavior." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16_sliding_window_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16_sliding_window_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3-flash-attention-sliding-window-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, and cache writes are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar and capped by the 2047-token sliding window." }, "evidence": [ { "label": "Phi-3 Mini 4K Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-mini-4k-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA f39ac1d28e925b323eae81227eaba4464caced4e, the API records a public/non-gated MIT text-generation repo with phi3, custom_code, English, French, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 578189. The API safetensors block reports BF16: 3821079552 and total: 3821079552." }, { "label": "Phi-3 Mini 4K Instruct config", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 4096 max position embeddings, original 4096 context, sliding_window 2047, and vocab size 32064." }, { "label": "Phi-3 Mini 4K Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/modeling_phi3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "lm_head_layout" ], "notes": "Manual review found Phi3Attention builds qkv_proj with output size num_heads*head_dim plus two num_key_value_heads*head_dim streams, then caches key_states and value_states. Phi3FlashAttention2 slices cached K/V tensors when kv_seq_len exceeds config.sliding_window and calls FlashAttention with window_size set from config.sliding_window. Phi3ForCausalLM instantiates a separate lm_head linear layer." }, { "label": "Phi-3 Mini 4K Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/f39ac1d28e925b323eae81227eaba4464caced4e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 7642159104 bytes across two shards. Range-read shard headers found 195 BF16 tensors totaling 3821079552 parameters / 7.642159104 GB. model.layers.* tensors total 3.624075264B parameters / 7.248150528 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, lm_head.weight is 98.500608M parameters / 0.197001216 GB, and model.embed_tokens.weight is 98.500608M parameters / 0.197001216 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 3.722578944B parameters / 7.445157888 GB. The resident-only subset is model.embed_tokens.weight: 0.098500608B parameters / 0.197001216 GB." }, { "label": "Phi-3 Mini 4K Instruct license", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights and applying the repo's 2047-token sliding-window K/V cache behavior." }, { "id": "microsoft--phi-3-vision-128k-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-3-vision-128k-instruct", "title": "Microsoft Phi-3 Vision 128K Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-3 Vision 128K Instruct repo.", "model_family": "phi3-v-dense", "architecture": { "canonical_architecture_id": "phi-3-vision-128k", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.14662144, "swept_params_b": 3.722578944, "auxiliary_resident_params_b": 0.424042496, "resident_weight_gb": 8.29324288, "swept_weight_gb": 7.445157888, "auxiliary_resident_weight_gb": 0.848084992, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and resident vision embedding/CLIP tensors while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.vision_embed_tokens.* and model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary generated text token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 96, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 32 attention heads, 32 KV heads, hidden size 3072, and max context 131072. The remote code builds qkv_proj with hidden_size to num_heads*head_dim plus two KV streams, then caches key_states and value_states. Config sliding_window equals max_position_embeddings, so it does not reduce the supported-context allocation/read horizon." }, "notes": "Phi3VForCausalLM is multimodal with a resident CLIP vision embedding path. This profile models ordinary text decode after any image/video prefill, not CLIP vision encoder or image projection throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3v-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, image projection, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-3 Vision 128K Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-vision-128k-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "weight_format", "total_params_b", "multimodal" ], "notes": "At repo SHA ed2772fabe9dc9acd0caad54b62761d92520cc44, the API records a public non-gated MIT text-generation repo with transformers, safetensors, phi3_v, custom_code, vision, multilingual, region:us, and 250174 downloads. The API safetensors block reports BF16: 4146621440 and total: 4146621440. The card describes Phi-3 Vision 128K Instruct as a lightweight multimodal model with 128K context support." }, { "label": "Phi-3 Vision 128K Instruct config", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Phi3VForCausalLM, phi3_v, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 3072, intermediate size 8192, 32 attention heads, 32 KV heads, 131072 max position embeddings, original 4096 context, SU rope scaling, sliding_window 131072, vocab size 32064, CLIP vision model openai/clip-vit-large-patch14-336, image_dim_out 1024, and image MLP projection settings." }, { "label": "Phi-3 Vision 128K Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/modeling_phi3_v.py", "source_type": "manual_review", "supports": [ "kv_adapter", "vision_embedding_scope", "lm_head_layout", "sliding_window_boundary" ], "notes": "Manual review found Phi3Attention builds qkv_proj with output size num_heads*head_dim + 2*num_key_value_heads*head_dim and caches key_states/value_states through past_key_value.update. Phi3VModel instantiates model.vision_embed_tokens from Phi3ImageEmbedding when embd_layer is present, and Phi3VForCausalLM instantiates a separate lm_head linear layer. The FlashAttention sliding-window path only slices cache when kv_seq_len exceeds config.sliding_window." }, { "label": "Phi-3 Vision 128K Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_resident_scope" ], "notes": "The index records metadata total_size 8293242880 bytes across two shards. Range-read shard headers found 592 BF16 tensors totaling 4146621440 parameters / 8.293242880 GB. model.layers.* tensors total 3.624075264B parameters / 7.248150528 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, lm_head.weight is 98.500608M parameters / 0.197001216 GB, model.embed_tokens.weight is 98.500608M parameters / 0.197001216 GB, and model.vision_embed_tokens.* tensors total 325.541888M parameters / 0.651083776 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 3.722578944B parameters / 7.445157888 GB. The resident-only subset is model.embed_tokens.weight plus model.vision_embed_tokens.*: 0.424042496B parameters / 0.848084992 GB." }, { "label": "Phi-3 Vision 128K Instruct license", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only vision and input-embedding tensors from per-token swept text/logit weights." }, { "id": "microsoft--phi-4-mini-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-4-mini-instruct", "title": "Microsoft Phi-4 Mini Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-4 Mini Instruct repo.", "model_family": "phi4-mini-dense", "architecture": { "canonical_architecture_id": "phi-4-mini", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.83602176, "swept_params_b": 3.83602176, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.67204352, "swept_weight_gb": 7.67204352, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode charges all stored BF16 tensors because input and output embeddings are tied and no separate lm_head.weight is stored", "auxiliary_scope": "no resident-only tensor subset is excluded for ordinary text decode", "notes": "The config records tie_word_embeddings true. The remote Phi3ForCausalLM class creates lm_head and marks it as a tied weight, while the safetensors headers store only model.embed_tokens.weight and no lm_head.weight. The tied embedding matrix is used as the output projection and is swept for logits in ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 24 attention heads, 8 KV heads, hidden size 3072, head_dim 128 by hidden/head ratio, 131072 max position embeddings, and sliding_window 262144. The remote Phi3 attention path builds separate K and V streams from qkv_proj. Because the sliding window is larger than the served 128K context, K/V residency and read traffic are full-context within the supported context." }, "notes": "Phi3ForCausalLM is used as the implementation class for Phi-4 Mini. The model card describes Phi-4-mini-instruct as a 3.8B dense decoder-only Transformer with 128K context. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-4 Mini Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-4-mini-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA cfbefacb99257ffa30c83adab238a50856ac3083, the API records a public/non-gated MIT text-generation repo with phi3, custom_code, multilingual, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 513426. The API safetensors block reports BF16: 3836021760 and total: 3836021760. The card describes Phi-4-mini-instruct as a 3.8B dense decoder-only Transformer with 128K context and MIT licensing." }, { "label": "Phi-4 Mini Instruct config", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/cfbefacb99257ffa30c83adab238a50856ac3083/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, torch_dtype bfloat16, tie_word_embeddings true, 32 decoder layers, hidden size 3072, intermediate size 8192, 24 attention heads, 8 KV heads, 131072 max position embeddings, original 4096 context, sliding_window 262144, partial_rotary_factor 0.75, and vocab size 200064." }, { "label": "Phi-4 Mini Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/cfbefacb99257ffa30c83adab238a50856ac3083/modeling_phi3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "lm_head_layout", "embedding_layout" ], "notes": "Manual review found Phi3Attention building qkv_proj with output size num_attention_heads*head_dim plus two num_key_value_heads*head_dim streams, then caching key_states and value_states. The causal mask honors config.sliding_window when set, but this repo's 262144-token window is larger than the 131072-token served context. Phi3ForCausalLM instantiates lm_head, declares _tied_weights_keys ['lm_head.weight'], exposes get_output_embeddings, and computes logits through lm_head." }, { "label": "Phi-4 Mini Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/resolve/cfbefacb99257ffa30c83adab238a50856ac3083/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 7672043520 bytes across two shards. Range-read shard headers found 194 BF16 tensors totaling 3836021760 parameters / 7.672043520 GB. model.layers.* tensors total 3221422080 parameters / 6.442844160 GB, model.norm.weight is 3072 parameters / 0.000006144 GB, and model.embed_tokens.weight is 614596608 parameters / 1.229193216 GB. There is no lm_head.weight tensor because output embeddings are tied. The swept ordinary text subset is all stored tensors: model.layers.*, model.norm.weight, and model.embed_tokens.weight." }, { "label": "Phi-4 Mini Instruct license", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/cfbefacb99257ffa30c83adab238a50856ac3083/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by charging tied input/output embedding bytes as swept output-projection traffic and treating the 262144-token sliding window as full-context within the supported 128K context." }, { "id": "microsoft--phi-4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/phi-4", "title": "Microsoft Phi-4 BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi-4 repo.", "model_family": "phi4-dense", "architecture": { "canonical_architecture_id": "phi-4", "max_context_tokens": 16384, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.6595072, "swept_params_b": 14.14570496, "auxiliary_resident_params_b": 0.51380224, "resident_weight_gb": 29.3190144, "swept_weight_gb": 28.29140992, "auxiliary_resident_weight_gb": 1.02760448, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup while including model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "The config records tie_word_embeddings false and the safetensors headers contain a separate lm_head.weight, so lm_head.weight is the swept output projection while model.embed_tokens.weight is resident-only for ordinary text decode. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 10, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 40 layers, 40 attention heads, 10 KV heads, hidden size 5120, head_dim 128, 16384 max position embeddings, and sliding_window null. The official Transformers Phi3 implementation builds qkv_proj with hidden_size to num_heads*head_dim plus two KV streams, then caches key_states and value_states." }, "notes": "Phi3ForCausalLM is used as the implementation class for Phi-4. The model card describes Phi-4 as a 14B dense decoder-only Transformer with 16K context. This profile models ordinary autoregressive text decode." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phi3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, and FlashAttention efficiency are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Phi-4 model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/phi-4", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA 932b33c0ec9ca189badeb22480721a8de9d0e006, the API records a public/non-gated MIT text-generation repo with phi3, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 869780. The API safetensors block reports BF16: 14659507200 and total: 14659507200. The card describes Phi-4 as a 14B dense decoder-only Transformer with 16K context and MIT licensing." }, { "label": "Phi-4 config", "url": "https://huggingface.co/microsoft/phi-4/raw/932b33c0ec9ca189badeb22480721a8de9d0e006/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Phi3ForCausalLM, model_type phi3, torch_dtype bfloat16, tie_word_embeddings false, 40 decoder layers, hidden size 5120, intermediate size 17920, 40 attention heads, 10 KV heads, 16384 max position embeddings, original 16384 context, rope_theta 250000, sliding_window null, and vocab size 100352." }, { "label": "Transformers 4.47.0 Phi3 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.47.0/src/transformers/models/phi3/modeling_phi3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "lm_head_layout", "embedding_layout", "sliding_window_boundary" ], "notes": "Manual review found Phi3Attention building qkv_proj with output size num_heads*head_dim plus two num_key_value_heads*head_dim streams, caching key_states and value_states, and using config.sliding_window only when it is not null. Phi3Model instantiates model.embed_tokens, and Phi3ForCausalLM instantiates a separate lm_head linear layer used for logits." }, { "label": "Phi-4 safetensors index and shard headers", "url": "https://huggingface.co/microsoft/phi-4/resolve/932b33c0ec9ca189badeb22480721a8de9d0e006/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records metadata total_size 29319014400 bytes across six shards. Range-read shard headers found 243 BF16 tensors totaling 14.659507200B parameters / 29.319014400 GB. model.layers.* tensors total 13.631897600B parameters / 27.263795200 GB, model.norm.weight is 5120 parameters / 0.000010240 GB, lm_head.weight is 513.802240M parameters / 1.027604480 GB, and model.embed_tokens.weight is 513.802240M parameters / 1.027604480 GB. The swept ordinary text subset is model.layers.* plus model.norm.weight plus lm_head.weight: 14.145704960B parameters / 28.291409920 GB. The resident-only subset is model.embed_tokens.weight: 0.513802240B parameters / 1.027604480 GB." }, { "label": "Phi-4 license", "url": "https://huggingface.co/microsoft/phi-4/raw/932b33c0ec9ca189badeb22480721a8de9d0e006/LICENSE", "source_type": "model_card", "supports": [ "license" ], "notes": "The model card metadata records MIT licensing and links to the repo LICENSE file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, official Transformers 4.47.0 Phi3 implementation, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights." }, { "id": "microsoft--phi-mini-moe-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-mini-MoE-instruct", "title": "Microsoft Phi Mini MoE Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi mini MoE Instruct repo.", "model_family": "phimoe-slimmoe", "architecture": { "canonical_architecture_id": "phi-mini-moe", "max_context_tokens": 4096, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 15.295265408, "main_resident_weight_gb": 15.03259712, "auxiliary_resident_weight_gb": 0.262668288, "fixed_weight_gb": 2.9530016, "routed_expert_weight_gb": 0.75497472, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, block_sparse_moe.gate, and lm_head, excluding resident-only input embedding", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "notes": "Header-derived stored bytes are used. Each routed expert stores w1, w2, and w3 BF16 tensors in each of 32 decoder layers. Routed expert bytes are exactly uniform across 16 expert indexes; router gate tensors are fixed always-on MoE traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 32, "kv_heads": 8, "head_dim": 128, "window_tokens": 2047, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window 2047. The remote FlashAttention path slices cached K/V tensors when kv_seq_len exceeds config.sliding_window, so the profile caps allocation and read traffic at 2047 live K/V tokens per layer." } ], "notes": "All 32 layers use the same sliding-window K/V geometry." }, "notes": "PhiMoEForCausalLM is a text-only sparse MoE decoder. This profile models ordinary autoregressive text decode with the FlashAttention sliding-window cache behavior documented by the repo's remote code." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16_sliding_window_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16_sliding_window_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phimoe-flash-attention-sliding-window-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, router compute, expert compute, and cache writes are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar and capped by the 2047-token sliding window." }, "evidence": [ { "label": "Phi mini MoE Instruct API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-mini-MoE-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At repo SHA f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd, the API records a public non-gated MIT text-generation repo with transformers, safetensors, phimoe, custom_code, en, and region:us tags. Current downloads are 196717. The API safetensors block reports BF16 7647632704 and total 7647632704." }, { "label": "Phi mini MoE Instruct model card", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct", "source_type": "model_card", "supports": [ "architecture", "active_params_b", "max_context_tokens", "training_lineage" ], "notes": "The model card describes Phi-mini-MoE as a SlimMoE model with 7.6B total parameters, 2.4B activated parameters, and 4K context. It also states that the model was compressed and distilled from the base model shared by Phi-3.5-MoE and GRIN-MoE and post-trained for instruction following and safety." }, { "label": "Phi mini MoE Instruct config", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/raw/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records PhiMoEForCausalLM, model_type phimoe, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 4096, intermediate size 960, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 4096, sliding_window 2047, vocab size 32064, attention_bias true, lm_head_bias true, 16 local experts, and 2 experts per token." }, { "label": "Phi mini MoE Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/raw/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/modeling_slimmoe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "routing_model", "expert_layout", "lm_head_layout" ], "notes": "Manual review found the mini repo's modeling_slimmoe.py has SHA-256 9ab50e875106cbf6544db2a231727bfe0fdb897776c0875179573e03077c0c02, matching the already audited Phi tiny MoE runtime code. PhiMoEAttention builds separate q_proj, k_proj, v_proj, and o_proj tensors, PhiMoEFlashAttention2 slices cached K/V tensors when kv_seq_len exceeds config.sliding_window, PhiMoESparseMoeBlock routes with sparsemixer top_k=2 through a per-layer gate and 16 local experts, each expert containing w1, w2, and w3, and PhiMoEForCausalLM instantiates a separate lm_head linear layer." }, { "label": "Phi mini MoE Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/resolve/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across all four indexed shards. Stored tensors sum to the index total_size, 15.295265408 GB, across 1957 BF16 tensors. Linked shard sizes sum to 15.295512816 GB, with 247408 bytes of safetensors header/container overhead outside tensor payloads. Ordinary text resident tensors, defined as model.* excluding model.embed_tokens.weight plus lm_head.weight and lm_head.bias, sum to 15.032597120 GB. model.embed_tokens.weight is the only resident-only auxiliary tensor at 0.262668288 GB. Routed expert tensors, defined as model.layers.*.block_sparse_moe.experts.* w1/w2/w3 weights, sum to 12.079595520 GB and divide exactly into 16 uniform expert indexes of 0.754974720 GB. Fixed ordinary text traffic, including attention tensors, router gates, norms, and lm_head tensors, sums to 2.953001600 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights, using exact routed-expert bytes, and applying the repo's sliding-window K/V cache behavior." }, { "id": "microsoft--phi-tiny-moe-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "microsoft/Phi-tiny-MoE-instruct", "title": "Microsoft Phi Tiny MoE Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Phi tiny MoE Instruct repo.", "model_family": "phimoe-slimmoe", "architecture": { "canonical_architecture_id": "phi-tiny-moe", "max_context_tokens": 4096, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 7.510440576, "main_resident_weight_gb": 7.247772288, "auxiliary_resident_weight_gb": 0.262668288, "fixed_weight_gb": 1.610627712, "routed_expert_weight_gb": 0.352321536, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, block_sparse_moe.gate, and lm_head, excluding resident-only input embedding", "auxiliary_scope": "model.embed_tokens.weight is resident for the checkpoint but not swept as a full matrix for each ordinary generated token", "notes": "Header-derived stored bytes are used. Each routed expert stores w1, w2, and w3 BF16 tensors in each of 32 decoder layers. Routed expert bytes are exactly uniform across 16 expert indexes; router gate tensors are fixed always-on MoE traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 32, "kv_heads": 4, "head_dim": 128, "window_tokens": 2047, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window 2047. The remote FlashAttention path slices cached K/V tensors to sliding_window - 1 before appending the next token, so the profile caps allocation and read traffic at 2047 live K/V tokens per layer." } ], "notes": "All 32 layers use the same sliding-window K/V geometry." }, "notes": "PhiMoEForCausalLM is a text-only sparse MoE decoder. This profile models ordinary autoregressive text decode with the FlashAttention sliding-window cache behavior documented by the repo's remote code." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16_sliding_window_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16_sliding_window_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-phimoe-flash-attention-sliding-window-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, router compute, expert compute, and cache writes are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. KV cache is charged at BF16 two bytes per scalar and capped by the 2047-token sliding window." }, "evidence": [ { "label": "Phi tiny MoE Instruct model card and API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-tiny-MoE-instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At repo SHA 2fe50e88d0e2a5a132563815686ea0dcc8e252b5, the API records a public/non-gated MIT text-generation repo with transformers, safetensors, phimoe, custom_code, en, and region:us tags. Current downloads are 884992. The API safetensors block reports BF16: 3755220288 and total: 3755220288. The model card describes Phi-tiny-MoE as a 3.8B total / 1.1B activated parameter SlimMoE model, with 16 experts, 2 activated experts, and 4K context." }, { "label": "Phi tiny MoE Instruct config", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/raw/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records PhiMoEForCausalLM, model_type phimoe, torch_dtype bfloat16, tie_word_embeddings false, 32 decoder layers, hidden size 4096, intermediate size 448, 16 attention heads, 4 KV heads, head_dim 128, max_position_embeddings 4096, sliding_window 2047, vocab size 32064, attention_bias true, lm_head_bias true, 16 local experts, and 2 experts per token." }, { "label": "Phi tiny MoE Instruct remote modeling code", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/raw/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/modeling_slimmoe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "sliding_window", "routing_model", "expert_layout", "lm_head_layout" ], "notes": "Manual review found PhiMoEAttention building separate q_proj, k_proj, v_proj, and o_proj tensors, caching key_states and value_states, and PhiMoEFlashAttention2 slicing cached K/V tensors when kv_seq_len exceeds config.sliding_window. PhiMoESparseMoeBlock routes with sparsemixer top_k=2 through a per-layer gate and 16 local experts, each expert containing w1, w2, and w3. PhiMoEForCausalLM instantiates a separate lm_head linear layer." }, { "label": "Phi tiny MoE Instruct safetensors index and shard headers", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/resolve/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across both indexed shards. Stored tensors sum to the index total_size, 7.510440576 GB, across 1957 BF16 tensors. Ordinary text resident tensors, defined as model.* excluding model.embed_tokens.weight plus lm_head.weight and lm_head.bias, sum to 7.247772288 GB. model.embed_tokens.weight is the only resident-only auxiliary tensor at 0.262668288 GB. Routed expert tensors, defined as model.layers.*.block_sparse_moe.experts.* w1/w2/w3 weights, sum to 5.637144576 GB and divide exactly into 16 uniform expert indexes of 0.352321536 GB. Fixed ordinary text traffic, including attention tensors, router gates, norms, and lm_head tensors, sums to 1.610627712 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, remote modeling code, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by separating resident-only input embedding bytes from per-token swept text/logit weights, using exact routed-expert bytes, and applying the repo's sliding-window K/V cache behavior." }, { "id": "minimaxai--minimax-m2-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MiniMaxAI/MiniMax-M2.5", "title": "MiniMax M2.5 FP8", "summary": "Audited memory-side bounds profile for the FP8 MiniMax M2.5 MoE repo.", "model_family": "minimax-m2-moe", "architecture": { "canonical_architecture_id": "minimax-m2-5", "max_context_tokens": 196608, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 230.12163072, "main_resident_weight_gb": 228.892437504, "auxiliary_resident_weight_gb": 1.229193216, "fixed_weight_gb": 4.1571072, "routed_expert_weight_gb": 0.877872384, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_data_offsets_fp8_bf16_f32", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header data_offsets are used as the byte source of truth because model.safetensors.index.json metadata.total_size disagrees with the mapped tensor payload and HF usedStorage. Expert tensors are stored as per-expert w1/w2/w3 matrices plus scale tensors and divide exactly across 256 expert indexes. The config records shared_intermediate_size 0, so there are no shared expert tensors. The config enables MTP, but the checkpoint headers contain no MTP-named tensors, so no separate MTP sidecar traffic is charged." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config omits sliding_window, and MiniMaxM2Config defaults sliding_window to null. The custom model code chooses create_causal_mask when sliding_window is null and updates standard past_key_values with K and V tensors, so this profile charges full-context K and V streams for all layers." }, "notes": "MiniMaxM2ForCausalLM profile using the served custom config, model code, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-minimax-m2-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored F8_E4M3, BF16, and F32 safetensors bytes. FP8 dequantization, router compute, expert compute, and writes are outside this memory-side bound.", "notes": "The config records FP8 quantization with dynamic activations, E4M3 weights, 128x128 weight blocks, and gate/e_score_correction_bias/lm_head left unconverted. No repo evidence was found for quantized KV cache, so KV cache is charged as expanded BF16 K/V streams." }, "evidence": [ { "label": "MiniMax M2.5 model card and API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit f710177d938eff80b684d42c5aa84b382612f21f, the API records a public non-gated text-generation repo with transformers, custom_code, minimax_m2, fp8, endpoints_compatible, deploy:azure, region:us, license other / modified-mit, and safetensors tags. Current downloads were 698560 when audited. The API safetensors block reports tensor elements split across F32: 62654720, BF16: 1230021632, F8_E4M3: 227410968576, total 228703644928." }, { "label": "MiniMax M2.5 config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/f710177d938eff80b684d42c5aa84b382612f21f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records MiniMaxM2ForCausalLM, minimax_m2, hidden size 3072, intermediate size 1536, 62 layers, 48 attention heads, 8 KV heads, head_dim 128, 256 local experts, 8 experts per token, no shared expert, max_position_embeddings 196608, rope_theta 5000000, sigmoid routing with routing bias, use_qk_norm true, and FP8 quantization with gate, e_score_correction_bias, and lm_head excluded from conversion." }, { "label": "MiniMax M2.5 custom configuration and model code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/f710177d938eff80b684d42c5aa84b382612f21f/modeling_minimax_m2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "sliding_window" ], "notes": "Manual review found MiniMaxM2Attention updating standard past_key_values with K and V tensors, MiniMaxM2Model selecting create_causal_mask when config.sliding_window is null, and the sparse MoE block routing with sigmoid scores plus e_score_correction_bias and top-k equal to config.num_experts_per_tok. configuration_minimax_m2.py defaults sliding_window to null." }, { "label": "MiniMax M2.5 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/f710177d938eff80b684d42c5aa84b382612f21f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Range-read safetensors headers found 96103 mapped tensors across 125 shards. Summing tensor data_offsets gives 230.12163072 GB: 227.410968576 GB F8_E4M3, 2.460043264 GB BF16, and 0.25061888 GB F32. The index metadata.total_size field says 466.083044352 GB, but that conflicts with mapped tensor offsets and HF usedStorage, so it is not used for resident traffic. model.embed_tokens.weight has shape [200064, 3072] and contributes 1.229193216 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Routed expert tensors sum to 224.735330304 GB and divide exactly into 256 uniform expert groups of 0.877872384 GB. Non-expert fixed decode tensors including lm_head.weight sum to 4.1571072 GB. The headers contain no MTP-named tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, custom configuration/model code, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MiniMax M2 MoE FP8 profile for production profile-backed bounds." }, { "id": "minimaxai--minimax-m2-7", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MiniMaxAI/MiniMax-M2.7", "title": "MiniMax M2.7 FP8", "summary": "Audited memory-side bounds profile for the FP8 MiniMax M2.7 MoE repo.", "model_family": "minimax-m2-moe", "architecture": { "canonical_architecture_id": "minimax-m2-7", "max_context_tokens": 204800, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 230.12163072, "main_resident_weight_gb": 228.892437504, "auxiliary_resident_weight_gb": 1.229193216, "fixed_weight_gb": 4.1571072, "routed_expert_weight_gb": 0.877872384, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_mapped_tensor_offsets", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header data_offsets are used as the byte source of truth because model.safetensors.index.json metadata.total_size disagrees with the mapped tensor payload. Expert tensors are stored as per-expert w1/w2/w3 matrices plus scale tensors and divide exactly across 256 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config omits sliding_window, and the MiniMaxM2Config default is null. The model code chooses create_causal_mask when sliding_window is null, so this profile charges full-context K and V streams for all layers." }, "notes": "MiniMaxM2ForCausalLM profile using the served custom config, model code, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-minimax-m2-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored F8_E4M3, BF16, and F32 safetensors bytes. FP8 dequantization, router compute, expert compute, and writes are outside this memory-side bound.", "notes": "The config records dtype bfloat16 and FP8 quantization with dynamic activations, E4M3 weights, 128x128 weight blocks, and gate/e_score_correction_bias/lm_head left unconverted. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "MiniMax M2.7 model card and API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.7", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit d494266a4affc0d2995ba1fa35c8481cbd84294b, the API records a text-generation repo with custom_code, minimax_m2, fp8, and safetensors tags. The API safetensors block reports tensor elements split across F32: 62654720, BF16: 1230021632, F8_E4M3: 227410968576, total 228703644928." }, { "label": "MiniMax M2.7 config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/d494266a4affc0d2995ba1fa35c8481cbd84294b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records MiniMaxM2ForCausalLM, minimax_m2, bfloat16 dtype, hidden size 3072, intermediate size 1536, 62 layers, 48 attention heads, 8 KV heads, head_dim 128, 256 local experts, 8 experts per token, no shared expert, max_position_embeddings 204800, rope_theta 5000000, sigmoid routing with routing bias, use_qk_norm true, and FP8 quantization with gate, e_score_correction_bias, and lm_head excluded from conversion." }, { "label": "MiniMax M2.7 custom model code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/d494266a4affc0d2995ba1fa35c8481cbd84294b/modeling_minimax_m2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token" ], "notes": "Manual review found MiniMaxM2Attention updating standard past_key_values with K and V tensors and MiniMaxM2Model selecting create_causal_mask when config.sliding_window is null. The sparse MoE block routes with sigmoid scores plus e_score_correction_bias and top-k equal to config.num_experts_per_tok." }, { "label": "MiniMax M2.7 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/d494266a4affc0d2995ba1fa35c8481cbd84294b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Range-read safetensors headers found 96103 mapped tensors across 125 shards. Summing tensor data_offsets gives 230.12163072 GB: 227.410968576 GB F8_E4M3, 2.460043264 GB BF16, and 0.25061888 GB F32. The index metadata.total_size field says 480.836588544 GB, but that conflicts with mapped tensor offsets and per-file HEAD checks, so it is not used for resident traffic. model.embed_tokens.weight has shape [200064, 3072] and contributes 1.229193216 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Routed expert tensors sum to 224.735330304 GB and divide exactly into 256 uniform expert groups of 0.877872384 GB. Non-expert fixed decode tensors including lm_head.weight sum to 4.1571072 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served config, custom configuration/model code, safetensors index, range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained MiniMax M2 MoE FP8 profile for production profile-backed bounds." }, { "id": "minimaxai--minimax-m2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MiniMaxAI/MiniMax-M2", "title": "MiniMax M2 FP8", "summary": "Audited memory-side bounds profile for the FP8 MiniMax M2 MoE repo.", "model_family": "minimax-m2-moe", "architecture": { "canonical_architecture_id": "minimax-m2", "max_context_tokens": 196608, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 230.12163072, "main_resident_weight_gb": 228.892437504, "auxiliary_resident_weight_gb": 1.229193216, "fixed_weight_gb": 4.1571072, "routed_expert_weight_gb": 0.877872384, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_data_offsets_fp8_bf16_f32", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header data_offsets are used as the byte source of truth because model.safetensors.index.json metadata.total_size disagrees with the mapped tensor payload and file content-range totals. Expert tensors are stored as per-expert w1/w2/w3 matrices plus scale tensors and divide exactly across 256 expert indexes. The config records shared_intermediate_size 0, so there are no shared expert tensors. The config enables MTP, but the checkpoint headers contain no MTP-named tensors, so no separate MTP sidecar traffic is charged." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null. The custom model code chooses create_causal_mask when sliding_window is null and updates standard past_key_values with K and V tensors, so this profile charges full-context K and V streams for all layers." }, "notes": "MiniMaxM2ForCausalLM profile using the served custom config, model code, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-minimax-m2-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored F8_E4M3, BF16, and F32 safetensors bytes. FP8 dequantization, router compute, expert compute, and writes are outside this memory-side bound.", "notes": "The config records FP8 quantization with dynamic activations, E4M3 weights, 128x128 weight blocks, and gate/e_score_correction_bias/lm_head left unconverted. No repo evidence was found for quantized KV cache, so KV cache is charged as expanded BF16 K/V streams." }, "evidence": [ { "label": "MiniMax M2 model card and API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit 757303d492a50514c312788b5247a4f696a4c6a3, the API records a public non-gated text-generation repo with transformers, custom_code, minimax_m2, fp8, endpoints_compatible, deploy:azure, region:us, license other / modified-mit, and safetensors tags. Current downloads were 113607 when audited. The API safetensors block reports tensor elements split across F32: 62654720, BF16: 1230021632, F8_E4M3: 227410968576, total 228703644928. The model card describes MiniMax-M2 as a 230B-total / 10B-active MoE model for coding and agentic workflows." }, { "label": "MiniMax M2 config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records MiniMaxM2ForCausalLM, minimax_m2, hidden size 3072, intermediate size 1536, 62 layers, 48 attention heads, 8 KV heads, head_dim 128, 256 local experts, 8 experts per token, no shared expert, max_position_embeddings 196608, rope_theta 5000000, sigmoid routing with routing bias, use_qk_norm true, tie_word_embeddings false, use_mtp true with mtp_num_layers null, and FP8 quantization with gate, e_score_correction_bias, and lm_head excluded from conversion." }, { "label": "MiniMax M2 custom configuration and model code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/modeling_minimax_m2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "sliding_window" ], "notes": "Manual review found MiniMaxM2Attention updating standard past_key_values with K and V tensors, MiniMaxM2Model selecting create_causal_mask when config.sliding_window is null, and the sparse MoE block routing with sigmoid scores plus e_score_correction_bias and top-k equal to config.num_experts_per_tok. configuration_minimax_m2.py defaults sliding_window to null." }, { "label": "MiniMax M2 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Range-read safetensors headers found 96103 mapped tensors across 125 shards. The combined linked file size from content-range reads is 230.134264592 GB, with 0.012633872 GB of safetensors header bytes and 230.12163072 GB of tensor payload. Payload dtypes are 227.410968576 GB F8_E4M3, 2.460043264 GB BF16, and 0.25061888 GB F32. The index metadata.total_size field says 480.836588544 GB, but that conflicts with mapped tensor offsets and linked file sizes, so it is not used for resident traffic. model.embed_tokens.weight has shape [200064, 3072] and contributes 1.229193216 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Routed expert tensors sum to 224.735330304 GB and divide exactly into 256 uniform expert groups of 0.877872384 GB. Non-expert fixed decode tensors including lm_head.weight sum to 4.1571072 GB. The headers contain no MTP-named tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served config, custom configuration/model code, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MiniMax M2 MoE FP8 profile for production profile-backed bounds." }, { "id": "minimaxai--minimax-m3-mxfp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "MiniMaxAI/MiniMax-M3-MXFP8", "title": "MiniMax M3 MXFP8", "summary": "Unsupported profile stub with exact resident tensor evidence for the MXFP8 MiniMax M3 multimodal sparse-attention MoE repo.", "model_family": "minimax-m3-vl-sparse-moe", "base_model_proof": { "base_model": "MiniMaxAI/MiniMax-M3", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served MXFP8 config, base BF16 config comparison, and safetensors header layout", "config_compatible": false, "notes": "The repo records MiniMaxAI/MiniMax-M3 as its quantized base and preserves the same top-level architecture and vision config. The served MXFP8 text_config changes dtype and MTP metadata relative to the BF16 base, so this profile uses the served MXFP8 config directly rather than inheriting the base wholesale." }, "architecture": { "canonical_architecture_id": "minimax-m3-vl", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 443.74208512, "main_resident_weight_gb": 439.549553664, "auxiliary_resident_weight_gb": 4.192531456, "fixed_weight_gb": 13.517319168, "routed_expert_weight_gb": 3.328376832, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_data_offsets_mxfp8_bf16_f32_u8", "traffic_scope": "Exact ordinary text weight groups are recorded for evidence only; Bounds Engine v1 does not use them for production throughput because MiniMax Sparse Attention has no audited KV/state adapter here.", "auxiliary_scope": "vision_tower, multi_modal_projector, patch_merge_mlp, and language_model.model.embed_tokens.weight are resident for the package but are not swept as full matrices for each ordinary generated text token.", "shared_expert_notes": "The config records n_shared_experts 1 and 4 routed experts per token. Header grouping keeps all always-on dense, shared, gate, attention, norm, and lm_head traffic in fixed_weight_gb while routed expert tensors are divided across 128 uniform expert indexes.", "notes": "Range-read safetensors headers record 443.742085120 GB across 45,838 tensors. Text-decode main resident tensors are the language package excluding input embeddings. Routed expert tensors in layers 3-59 total 426.032234496 GB and divide exactly into 128 expert groups of 3.328376832 GB each." }, "kv_adapter": { "kind": "unknown", "reason": "MiniMax M3 uses MiniMax Sparse Attention for million-token context. Bounds Engine v1 has no audited MSA allocation/read traffic formula, so using full-context KV or inventing a compressed ratio would produce misleading throughput.", "notes": "The served config records sparse_attention_config with sparse_block_size 128, sparse_topk_blocks 16, sparse_num_index_heads 4, sparse_index_dim 128, and sparse attention enabled on layers 3-59. Those fields identify the sparse-attention structure but are not enough by themselves to prove per-session allocation traffic or per-output-token read traffic." }, "notes": "MiniMaxM3SparseForConditionalGeneration is a multimodal VLM wrapper around a MiniMaxM3SparseForCausalLM text backbone. This profile intentionally fails closed until an audited MiniMax Sparse Attention KV/state adapter exists." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0078637243865287, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-minimax-m3-msa-mxfp8", "dequantization_notes": "The MXFP8 artifact stores most matrix payload in F8_E4M3 plus U8 scale tensors, with BF16/F32 ignored-layer tensors. Bounds Engine v1 records exact stored bytes but does not convert them into production tok/s because the sparse-attention state traffic is unsupported.", "notes": "The served quantization_config records quant_method mxfp8, dynamic activation quantization, weight_block_size [1, 32], and ignored BF16/F32 layers including lm_head, input embeddings, vision/projector tensors, and MoE gates." }, "evidence": [ { "label": "MiniMax M3 MXFP8 API metadata and model card", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3-MXFP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit ca165902e868fd015fab5acaff776643be00dc6e, the API records a public non-gated image-text-to-text repo with transformers, custom_code, minimax_m3_vl, multimodal, moe, arxiv:2606.13392, endpoints_compatible, mxfp8, region:us, license other / minimax-community, and 699,967 downloads. The API safetensors block reports F32 46,339,200; BF16 3,323,221,760; F8_E4M3 423,670,579,200; U8 13,239,705,600; total 440,279,845,760 logical elements." }, { "label": "MiniMax M3 MXFP8 served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8/raw/ca165902e868fd015fab5acaff776643be00dc6e/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "weight_format" ], "notes": "The config records MiniMaxM3SparseForConditionalGeneration / MiniMaxM3SparseForCausalLM, dtype bfloat16, 60 text layers, hidden size 6144, 64 attention heads, 4 KV heads, head_dim 128, vocab size 200064, untied embeddings, 1,048,576 max position embeddings, 128 local experts, 4 experts per token, 1 shared expert, MoE layers 3-59, and MXFP8 quantization. sparse_attention_config enables sparse attention on layers 3-59 with sparse_topk_blocks 16, sparse_block_size 128, sparse_num_index_heads 4, and sparse_index_dim 128." }, { "label": "MiniMax M3 MXFP8 model card sparse-attention notes", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8", "source_type": "model_card", "supports": [ "unsupported_reason", "max_context_tokens", "runtime_format" ], "notes": "The model card says MiniMax-M3 is a native multimodal model with 1M context, about 428B parameters, about 23B activated parameters, and MiniMax Sparse Attention. It says MSA improves long-context efficiency and gives 9x prefill and 15x decode speedups compared with M2 at 1M context. These speedups are not an allocation/read traffic formula for Bounds Engine v1." }, { "label": "MiniMax M3 BF16 base config comparison", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/fa6849585c885749b76c0b1f299841b231e4cbfa/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the MXFP8 repo and BF16 base share top-level architecture and vision_config. The text configs differ only in dtype and MTP metadata: MXFP8 adds dtype bfloat16 and num_mtp_modules 1, while the BF16 base records num_mtp_modules 7 and num_nextn_predict_layers 1. Core text geometry, sparse attention fields, MoE geometry, context, vocab, and tied-embedding fields match." }, { "label": "MiniMax M3 MXFP8 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8/raw/ca165902e868fd015fab5acaff776643be00dc6e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Range-read headers across all 31 shards found 45,838 tensors and 443.742085120 GB direct tensor payload: 423.670579200 GB F8_E4M3, 13.239705600 GB U8, 6.646443520 GB BF16, and 0.185356800 GB F32. model.safetensors.index.json metadata.total_size reports 451.543283200 GB, which conflicts with direct data_offsets and HF usedStorage, so direct header spans are used. language_model tensors total 442.007940096 GB, vision_tower tensors 1.265382400 GB, and projector/patch-merge tensors 0.468762624 GB. input embeddings and lm_head are separate 2.458386432 GB BF16 tensors. Routed expert tensors in layers 3-59 total 426.032234496 GB, exactly 3.328376832 GB per expert group across 128 experts. No MTP-named tensors were present." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from live HF API metadata, served MXFP8 config, BF16 base config comparison, model card sparse-attention notes, and direct safetensors shard-header range reads. Marked unsupported because Bounds Engine v1 lacks an audited MiniMax Sparse Attention KV/state adapter." }, "unsupported_reason": "MiniMax-M3-MXFP8 uses MiniMax Sparse Attention for long-context decode, and Bounds Engine v1 has no audited MSA allocation/read traffic formula. Production tok/s bounds are disabled rather than guessing full-context KV or a sparse compression ratio.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after MiniMax Sparse Attention has a dedicated KV/state adapter with verified allocation and read formulas." }, { "id": "minimaxai--minimax-m3", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "MiniMaxAI/MiniMax-M3", "title": "MiniMax M3", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 MiniMax M3 multimodal sparse-attention MoE repo.", "model_family": "minimax-m3-vl-sparse-moe", "base_model_proof": { "base_model": "MiniMaxAI/MiniMax-M3", "relation": "base", "source": "Hugging Face API metadata, served BF16 config, model card, and safetensors header layout", "config_compatible": true, "notes": "This is the base BF16/F32 MiniMax M3 repository. The served config records the full multimodal wrapper and text backbone directly, so no repo-name inference or quantized-child inheritance is used." }, "architecture": { "canonical_architecture_id": "minimax-m3-vl", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 854.17295872, "main_resident_weight_gb": 849.980427264, "auxiliary_resident_weight_gb": 4.192531456, "fixed_weight_gb": 23.736093696, "routed_expert_weight_gb": 6.455033856, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_data_offsets_bf16_f32", "traffic_scope": "Exact ordinary text weight groups are recorded for evidence only; Bounds Engine v1 does not use them for production throughput because MiniMax Sparse Attention has no audited KV/state adapter here.", "auxiliary_scope": "vision_tower, multi_modal_projector, patch_merge_mlp, and language_model.model.embed_tokens.weight are resident for the package but are not swept as full matrices for each ordinary generated text token.", "shared_expert_notes": "The config records n_shared_experts 1 and 4 routed experts per token. Header grouping keeps all always-on dense, shared, gate, attention, norm, and lm_head traffic in fixed_weight_gb while routed expert tensors are divided across 128 uniform expert indexes.", "notes": "Range-read safetensors headers record 854.172958720 GB across 23,416 tensors. Text-decode main resident tensors are the language package excluding input embeddings. Routed expert tensors in layers 3-59 total 826.244333568 GB and divide exactly into 128 expert groups of 6.455033856 GB each." }, "kv_adapter": { "kind": "unknown", "reason": "MiniMax M3 uses MiniMax Sparse Attention for million-token context. Bounds Engine v1 has no audited MSA allocation/read traffic formula, so using full-context KV or inventing a compressed ratio would produce misleading throughput.", "notes": "The served config records sparse_attention_config with sparse_block_size 128, sparse_topk_blocks 16, sparse_num_index_heads 4, sparse_index_dim 128, and sparse attention enabled on layers 3-59. Those fields identify the sparse-attention structure but are not enough by themselves to prove per-session allocation traffic or per-output-token read traffic." }, "notes": "MiniMaxM3SparseForConditionalGeneration is a multimodal VLM wrapper around a MiniMaxM3SparseForCausalLM text backbone. This profile intentionally fails closed until an audited MiniMax Sparse Attention KV/state adapter exists." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0002170250318043, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-minimax-m3-msa-bf16", "dequantization_notes": "The base artifact stores almost all tensor payload in BF16 with F32 side tensors. Bounds Engine v1 records exact stored bytes but does not convert them into production tok/s because the sparse-attention state traffic is unsupported.", "notes": "The repo is public and non-gated, and the served config records torch_dtype bfloat16. The profile uses direct safetensors header spans for stored bytes because model.safetensors.index.json metadata.total_size overstates the direct tensor payload." }, "evidence": [ { "label": "MiniMax M3 API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit fa6849585c885749b76c0b1f299841b231e4cbfa, the API records a public non-gated image-text-to-text repo with transformers, safetensors, custom_code, minimax_m3_vl, multimodal, moe, arxiv:2606.13392, endpoints_compatible, region:us, license other, and 226,766 downloads. The API safetensors block reports BF16 426,993,800,960; F32 46,339,200; total 427,040,140,160 logical elements." }, { "label": "MiniMax M3 served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/fa6849585c885749b76c0b1f299841b231e4cbfa/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "weight_format" ], "notes": "The config records MiniMaxM3SparseForConditionalGeneration / MiniMaxM3SparseForCausalLM, dtype bfloat16, 60 text layers, hidden size 6144, 64 attention heads, 4 KV heads, head_dim 128, vocab size 200064, untied embeddings, 1,048,576 max position embeddings, 128 local experts, 4 experts per token, 1 shared expert, MoE layers 3-59, num_mtp_modules 7, and num_nextn_predict_layers 1. sparse_attention_config enables sparse attention on layers 3-59 with sparse_topk_blocks 16, sparse_block_size 128, sparse_num_index_heads 4, and sparse_index_dim 128." }, { "label": "MiniMax M3 model card sparse-attention notes", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3", "source_type": "model_card", "supports": [ "unsupported_reason", "max_context_tokens", "runtime_format" ], "notes": "The model card says MiniMax-M3 is a native multimodal model with 1M context, about 428B parameters, about 23B activated parameters, and MiniMax Sparse Attention. It says MSA improves long-context efficiency and gives 9x prefill and 15x decode speedups compared with M2 at 1M context. These speedups are not an allocation/read traffic formula for Bounds Engine v1." }, { "label": "MiniMax M3 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/fa6849585c885749b76c0b1f299841b231e4cbfa/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Range-read headers across all 59 shards found 23,416 tensors and 854.172958720 GB direct tensor payload: 853.987601920 GB BF16 and 0.185356800 GB F32. model.safetensors.index.json metadata.total_size reports 869.157697024 GB, so direct data_offsets are used as the byte source. language_model tensors total 852.438813696 GB, vision_tower tensors 1.265382400 GB, projector tensors 0.091250688 GB, and patch-merge tensors 0.377511936 GB. input embeddings and lm_head are separate 2.458386432 GB BF16 tensors. Routed expert tensors in layers 3-59 total 826.244333568 GB, exactly 6.455033856 GB per expert group across 128 experts. No MTP-named tensors were present." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from live HF API metadata, served BF16 config, model card sparse-attention notes, and direct safetensors shard-header range reads. Marked unsupported because Bounds Engine v1 lacks an audited MiniMax Sparse Attention KV/state adapter." }, "unsupported_reason": "MiniMax-M3 uses MiniMax Sparse Attention for long-context decode, and Bounds Engine v1 has no audited MSA allocation/read traffic formula. Production tok/s bounds are disabled rather than guessing full-context KV or a sparse compression ratio.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after MiniMax Sparse Attention has a dedicated KV/state adapter with verified allocation and read formulas." }, { "id": "minimaxai--minimax-vl-01", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "MiniMaxAI/MiniMax-VL-01", "title": "MiniMax VL 01 BF16/F32", "summary": "Audited memory-side ordinary text-decode profile for the MiniMax-VL-01 multimodal MiniMaxText01 Lightning-Attention MoE repo.", "model_family": "minimax-vl-01-lightning-moe", "base_model_proof": { "base_model": "MiniMaxAI/MiniMax-VL-01", "relation": "base", "source": "Hugging Face API metadata, model card, served config, custom text/VL modeling code, and safetensors header layout", "config_compatible": true, "notes": "This is the base BF16/F32 MiniMax-VL-01 repository. The profile models ordinary text decode through the embedded MiniMaxText01 language backbone after multimodal prefill; it does not include image encoder/projector compute or the model-card int8 serving option." }, "architecture": { "canonical_architecture_id": "minimax-vl-01-text01", "max_context_tokens": 8192, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 915.366670336, "main_resident_weight_gb": 909.75238144, "auxiliary_resident_weight_gb": 5.614288896, "fixed_weight_gb": 40.021504, "routed_expert_weight_gb": 27.17908992, "routed_experts": 32, "routed_experts_per_token": 2, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_data_offsets_bf16_f32", "traffic_scope": "ordinary text decode through language_model excluding input embedding lookup, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "vision_tower, multi_modal_projector, image_newline, and language_model.model.embed_tokens.weight are resident for the multimodal package but are not swept as full matrices for each ordinary generated text token.", "shared_expert_notes": "The served config records shared_intermediate_size [0], so there is no shared MoE expert path. All layers route to two of 32 local experts.", "notes": "Range-read safetensors headers record 915.366670336 GB across 8,637 tensors. Text tensors total 914.669154304 GB, vision_tower tensors 0.609398784 GB, multi_modal_projector tensors 0.088104960 GB, and image_newline is 0.000012288 GB. language_model.model.embed_tokens.weight is 4.916772864 GB F32 and resident-only for ordinary decode. Fixed text traffic excluding input embeddings and routed experts is 40.021504000 GB: 35.233300480 GB Lightning-attention weights, 2.264924160 GB full-attention weights, 2.458386432 GB lm_head, 0.062914560 GB router gates, 0.001966080 GB layer norms, and 0.000012288 GB final norm. Routed expert tensors total 869.730877440 GB and divide exactly into 32 uniform expert groups of 27.179089920 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "recurrent_state", "alloc_gb_per_session": 0.29360128, "read_gb_per_output_token": 0.29360128, "state_formula": "70 Lightning-Attention layers * 64 heads * 128 key dimension * 128 value dimension * 4 bytes", "notes": "MiniMaxText01LightningAttention maintains a decayed per-layer KV matrix state. The code initializes kv as float32 with shape batch, heads, key_dim, value_dim, updates it recurrently, and returns that state as the layer cache." }, { "kind": "full_context", "layers": 10, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "attn_type_list marks layers 7, 15, 23, 31, 39, 47, 55, 63, 71, and 79 as full attention. Those layers use the standard MiniMaxText01FlashAttention2 path with cached K and V tensors." } ], "notes": "The served text config records 80 layers with attn_type_list containing 70 Lightning-Attention layers and 10 full-attention layers. This profile charges the Lightning state as fixed recurrent FP32 state and the full-attention layers as ordinary BF16 K/V cache." }, "notes": "MiniMaxVL01ForConditionalGeneration wraps a CLIP vision tower and a MiniMaxText01ForCausalLM language model. This profile is for ordinary language decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0054601850472866, "kv_store_format": "bf16-full-attention-and-fp32-lightning-state", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-full-attention-and-fp32-lightning-state", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-minimax-vl01-text01-lightning-moe-memory-bound", "dequantization_notes": "The model card recommends optional Quanto int8 loading for deployment, but this profile covers the stored public BF16/F32 safetensors artifact. Quantized runtime variants need separate profiles.", "notes": "Bounds Engine v1 charges stored BF16/F32 checkpoint bytes for weight traffic, BF16 scalar K/V traffic for the 10 full-attention layers, and explicit FP32 recurrent-state bytes for the 70 Lightning-Attention layers." }, "evidence": [ { "label": "MiniMax VL 01 API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-VL-01", "source_type": "derived_calculation", "supports": [ "repo", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 308b79934be140a43a0fb80f82b4e20d0ebe3cb8, the API reports a public non-gated image-text-to-text repo with safetensors, minimax_vl_01, custom_code, arxiv:2501.08313, region:us, and 112,771 downloads. The API safetensors block reports F32 1,246,115,840 parameters, BF16 455,191,103,488 parameters, and 456.437219328B total logical elements." }, { "label": "MiniMax VL 01 served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "routed_experts", "routed_experts_per_token", "kv_adapter", "vision_config" ], "notes": "The config records MiniMaxVL01ForConditionalGeneration wrapping MiniMaxText01ForCausalLM, torch_dtype bfloat16, 80 text layers, hidden size 6144, 64 attention heads, 8 KV heads, head_dim 128, 8192 max positions, 32 local experts, 2 experts per token, shared_intermediate_size [0], vocab size 200064, and a CLIP vision tower with hidden size 1024, 24 layers, 16 heads, 336 image size, and 14 patch size. attn_type_list contains 70 type-0 Lightning-Attention layers and 10 type-1 full-attention layers." }, { "label": "MiniMax VL 01 model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format", "vision_config" ], "notes": "The card says MiniMax-VL-01 uses a ViT-MLP-LLM framework with a 303M-parameter ViT, a two-layer MLP projector, and MiniMax-Text-01 as the base LLM. It recommends vLLM for production deployment and shows optional Quanto int8 loading while leaving vision_tower, image_newline, multi_modal_projector, lm_head, embed_tokens, coefficient, and router gates unconverted." }, { "label": "MiniMax VL 01 custom text modeling code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/modeling_minimax_text_01.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token" ], "notes": "Manual review found MiniMaxText01Model using attn_type_list to instantiate MiniMaxText01LightningAttention for type 0 and MiniMaxText01FlashAttention2 for type 1. The Lightning path stores a recurrent float32 kv matrix of shape batch, heads, key_dim, value_dim and updates it with a decay. The full-attention path updates ordinary cached K/V tensors. The sparse MoE block routes with softmax top-k equal to config.num_experts_per_tok across config.num_local_experts." }, { "label": "MiniMax VL 01 custom VL modeling code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/modeling_minimax_vl_01.py", "source_type": "manual_review", "supports": [ "base_model_proof", "vision_config", "ordinary_text_decode_scope" ], "notes": "Manual review found MiniMaxVL01ForConditionalGeneration constructing a CLIP vision tower, a MiniMaxVL01MultiModalProjector, image_newline, and a MiniMaxText01ForCausalLM language_model. Image features are merged into language-model embeddings before the language model forward path." }, { "label": "MiniMax VL 01 safetensors index and shard headers", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Range-read headers across all 414 shards found 8,637 mapped tensors and 915.366670336 GB direct tensor payload: 910.382206976 GB BF16 and 4.984463360 GB F32. model.safetensors.index.json metadata.total_size reports 912.873652224 GB, so direct data_offsets are used as the byte source. language_model tensors total 914.669154304 GB, vision_tower tensors 0.609398784 GB, multi_modal_projector tensors 0.088104960 GB, and image_newline is 0.000012288 GB. input embeddings and lm_head are separate: 4.916772864 GB F32 and 2.458386432 GB BF16. Routed expert tensors total 869.730877440 GB, exactly 27.179089920 GB per expert group across 32 experts." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from live HF API metadata, served config, model card, custom text/VL modeling code, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This profile is for the stored BF16/F32 MiniMax-VL-01 repo and ordinary text decode after any multimodal prefill. It does not profile the model-card int8 deployment recipe as a separate runtime artifact." }, { "id": "mlx-community--devstral-small-2-24b-instruct-2512-4bit", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit", "title": "MLX Devstral Small 2 24B Instruct 2512 4-bit", "summary": "Audited memory-side ordinary text-decode bounds profile for the MLX 4-bit Devstral Small 2 24B Instruct 2512 package.", "model_family": "devstral-small-2-24b-mistral3-dense", "base_model_proof": { "base_model": "mistralai/Devstral-Small-2-24B-Instruct-2512", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served MLX config comparison, params.json comparison, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The model card says this package was converted to MLX format from mistralai/Devstral-Small-2-24B-Instruct-2512 at revision c599e8e56f3f9110e97f0dc0450ce248e3334d84. Manual comparison found matching checked text and vision architecture fields between the MLX config/params files and the pinned base config/params files. The MLX repo adds 4-bit affine quantization metadata and stores packed U32 language weights plus BF16 scale/bias side tensors." }, "architecture": { "canonical_architecture_id": "devstral-small-2-24b-instruct-2512", "max_context_tokens": 393216, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 24.01136128, "swept_params_b": 22.90131456, "auxiliary_resident_params_b": 1.11004672, "resident_weight_gb": 15.10267904, "swept_weight_gb": 13.84727552, "auxiliary_resident_weight_gb": 1.25540352, "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16 for language, vision, and multimodal projector tensors", "swept_parameter_scope": "ordinary text decode charges language_model.model.layers.*, language_model.model.norm.weight, and language_model.lm_head.weight from the MLX package", "auxiliary_scope": "language_model.model.embed_tokens, vision_tower, and multi_modal_projector tensors are resident for ordinary multimodal serving but not swept for each generated text token after prefill", "notes": "Header-derived stored bytes are used because the MLX package stores packed U32 quantized language weights with BF16 scales and biases, while the vision tower, projector, norms, and lm_head use BF16 tensors. Logical parameter counts treat U32 weight elements as eight packed 4-bit values and exclude scale/bias side tensors from logical parameter counts while charging all stored side bytes in memory traffic." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text_config records 40 decoder layers, 8 KV heads, 128-dimensional key/value heads, max_position_embeddings 393216, and sliding_window null. The MLX config records weight quantization only and no KV cache quantization scheme." }, "notes": "This profile models ordinary text decode through the language model after any image prefill. Vision encoder and projector throughput are outside Bounds Engine v1, but their weights remain resident in the package footprint." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.6289805423310011, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-vlm-4bit-affine-mistral3-multimodal-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored MLX safetensors bytes: packed U32 language weights plus BF16 scales, biases, norms, embeddings, lm_head, vision tower, and projector tensors. Dequantization, activation traffic, vision prefill, Apple MLX scheduling overhead, and compute are outside this memory-side bound.", "notes": "The repo config records 4-bit affine MLX quantization with group_size 64 and mode affine. weight_bytes_per_param records resident stored bytes divided by reconstructed logical resident parameters for catalog display; exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "MLX Devstral Small 2 API metadata", "url": "https://huggingface.co/api/models/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 97228558fcbdd1adefd39d515155db3a894b37f7, the API reports a public non-gated Apache-2.0 MLX image-text-to-text repo with safetensors, mistral3, mistral-common, 4-bit, conversational, base_model mistralai/Devstral-Small-2-24B-Instruct-2512, region:us, 129826 downloads, and safetensors parameters BF16 1826114560, U32 2862612480, total 4688727040." }, { "label": "MLX Devstral Small 2 model card", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The card says this package was converted to MLX format from mistralai/Devstral-Small-2-24B-Instruct-2512 at revision c599e8e56f3f9110e97f0dc0450ce248e3334d84 using the current mlx-vlm checkout, and shows the mlx_vlm.generate path." }, { "label": "MLX Devstral Small 2 config", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "kv_adapter", "vision_scope", "serving" ], "notes": "The config records Mistral3ForConditionalGeneration, image-text-to-text multimodal metadata, text_config model_type ministral3, hidden size 5120, intermediate size 32768, 40 hidden layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 393216, sliding_window null, vocab_size 131072, untied embeddings, Pixtral vision_config with 24 layers and hidden size 1024, and top-level 4-bit affine MLX quantization with group_size 64." }, { "label": "MLX Devstral Small 2 params", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/params.json", "source_type": "config", "supports": [ "kv_adapter", "vision_scope", "max_context_tokens" ], "notes": "params.json records the same text geometry: dim 5120, 40 layers, head_dim 128, hidden_dim 32768, 32 attention heads, 8 KV heads, max_position_embeddings 393216, untied embeddings, and a Pixtral-style vision_encoder with 24 layers, 16 heads, hidden size 1024, patch size 14, and patch_merge projector." }, { "label": "Mistral Devstral Small 2 base config and params", "url": "https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512/raw/c599e8e56f3f9110e97f0dc0450ce248e3334d84/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching checked text and vision architecture fields between the MLX config/params files and the pinned base config/params files: Mistral3ForConditionalGeneration, ministral3 text geometry, 40 layers, 8 KV heads, 393216 max positions, null sliding window, untied embeddings, Pixtral vision tower, patch_merge projector settings, image token IDs, and rope/yarn fields. The base repo records FP8 quantization metadata, while the MLX repo stores 4-bit affine quantization metadata." }, { "label": "MLX Devstral Small 2 safetensors index and shard headers", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "vision_scope" ], "notes": "The index records total_size 15.102679040 GB across three safetensors shards and 1147 tensors. Range-reading all shard headers matched total_size exactly: U32 11.450449920 GB and BF16 3.652229120 GB. Linked shard sizes total 15.102831624 GB, leaving 0.000152584 GB of safetensors header/container overhead outside tensor payloads. language_model tensors total 14.224762880 GB, vision_tower tensors total 0.806610944 GB, and multi_modal_projector tensors total 0.071305216 GB. language_model.model.embed_tokens contributes 0.377487360 GB resident-only. Ordinary swept text-decode traffic, defined as language_model.model.layers.*, language_model.model.norm.weight, and language_model.lm_head.weight, totals 13.847275520 GB. Resident-only input-embedding plus vision/projector bytes total 1.255403520 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, the MLX model card, pinned served config and params files, pinned base config/params comparison, safetensors index metadata, linked shard HEAD checks, and direct range-read safetensors headers." }, "notes": "This profile targets the MLX community 4-bit Devstral Small 2 package specifically. It supersedes the generated catalog estimate, which referenced the older base model metadata, omitted the image-text-to-text pipeline, and treated the package as ideal flat 4-bit dense weights." }, { "id": "mlx-community--gpt-oss-20b-mxfp4-q8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "mlx-community/gpt-oss-20b-MXFP4-Q8", "title": "MLX gpt-oss 20B MXFP4 Q8", "summary": "Audited memory-side bounds profile for the MLX MXFP4/Q8 gpt-oss-20b package.", "model_family": "gpt-oss-moe", "base_model_proof": { "base_model": "openai/gpt-oss-20b", "relation": "quantized", "source": "Hugging Face metadata, MLX model card, served config, and config comparison against openai/gpt-oss-20b", "config_compatible": true, "notes": "The model card says this package was converted to MLX format from openai/gpt-oss-20b using mlx-lm 0.27.0. The served config matches the pinned OpenAI base config on checked architecture fields after excluding quantization metadata and Transformers version." }, "architecture": { "canonical_architecture_id": "gpt-oss-20b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 12.076119168, "main_resident_weight_gb": 11.460789888, "auxiliary_resident_weight_gb": 0.61532928, "fixed_weight_gb": 1.295173248, "routed_expert_weight_gb": 0.31767552, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_mxfp4_q8_u32_u8_bf16", "traffic_scope": "ordinary text decode through transformer layers and lm_head, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens weight, scales, and biases are resident for token lookup but ordinary decode does not sweep the full embedding matrices per generated token", "notes": "Header-derived bytes are used because the MLX package stores expert tensors as packed U32 weights with U8 scales and BF16 biases, while embeddings, attention, routers, and lm_head use MLX Q8 affine side tensors. Expert tensors are packed with 32 experts as a tensor dimension; dividing the packed expert payload by 32 gives the same per-expert byte group as the official OpenAI MXFP4 package." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 64, "notes": "The config has 12 full_attention layers." }, { "kind": "sliding_window", "layers": 12, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "notes": "The config has 12 sliding_attention layers with a 128-token sliding window." } ], "notes": "The served config alternates sliding_attention and full_attention across 24 layers, matching the official gpt-oss-20b architecture." }, "notes": "This is a derived MLX package of gpt-oss-20b. The logical model architecture is unchanged, but the resident and ordinary-decode traffic bytes differ from the official OpenAI safetensors artifact because the MLX package stores Q8 affine tensors for embeddings, attention, routers, and lm_head." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.5773970956794225, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-mxfp4-q8-moe-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored MLX quantized tensor bytes from safetensors headers. Dequantization and compute overhead are outside this memory-side bound.", "notes": "The repo API and safetensors headers report BF16, U32, and U8 tensors. The average stored bytes per API/index parameter is included for catalog display only; production bounds use the exact stored-byte adapter." }, "evidence": [ { "label": "MLX gpt-oss-20b MXFP4 Q8 API metadata", "url": "https://huggingface.co/api/models/mlx-community/gpt-oss-20b-MXFP4-Q8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model_proof", "serving", "downloads" ], "notes": "Current API metadata records a public non-gated Apache-2.0 MLX text-generation repo at commit 773a7da77e569019bb0fd17a554b263738d669a3 with 371761 downloads, region:us, base_model:openai/gpt-oss-20b, and safetensors parameters BF16 63142464, U32 2838159360, U8 1194393600, total 20914755648." }, { "label": "MLX gpt-oss-20b MXFP4 Q8 model card", "url": "https://huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The card says the model was converted to MLX format from openai/gpt-oss-20b using mlx-lm 0.27.0." }, { "label": "MLX gpt-oss-20b MXFP4 Q8 config", "url": "https://huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8/raw/773a7da77e569019bb0fd17a554b263738d669a3/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "max_context_tokens", "attention_pattern", "sliding_window", "kv_heads", "head_dim", "serving" ], "notes": "The config records GptOssForCausalLM, 24 hidden layers, 32 local experts, 4 experts per token, 131072 max position embeddings, alternating sliding/full attention, 128-token sliding windows, 8 KV heads, 64 head dimension, and MLX MXFP4/Q8 quantization settings." }, { "label": "MLX gpt-oss-20b MXFP4 Q8 safetensors index and range-read headers", "url": "https://huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8/raw/773a7da77e569019bb0fd17a554b263738d669a3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The index records total_size 12076119168 and total_parameters 20914755648. Range-read headers across the three shards found 775 tensors totaling 12.076119168 GB: BF16 0.126284928 GB, U32 11.352637440 GB, and U8 0.597196800 GB. Embedding tensors total 0.615329280 GB, lm_head tensors total 0.615329280 GB, packed expert tensors total 10.165616640 GB, and non-embedding non-expert ordinary-decode traffic totals 1.295173248 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current Hugging Face API metadata, the MLX model card, pinned served config, pinned OpenAI base config comparison, safetensors index metadata, linked shard HEAD checks, and direct range-read safetensors headers." }, "notes": "This profile targets the MLX community MXFP4/Q8 package specifically. It should not be reused for the official OpenAI MXFP4 artifact or for GGUF derivatives because their resident and traffic bytes differ." }, { "id": "mlx-community--kimi-k2-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "mlx-community/Kimi-K2.5", "title": "MLX Community Kimi K2.5 4-bit", "summary": "Audited memory-side ordinary text-decode bounds profile for the MLX 4-bit Kimi K2.5 package.", "model_family": "kimi-k2-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-K2.5", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served MLX config comparison, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The model card says this package was converted to MLX format from moonshotai/Kimi-K2.5 using mlx-lm 0.30.5. Manual comparison found matching checked text-architecture fields between the MLX config and the audited Moonshot base config. The MLX repo adds 4-bit affine quantization metadata and stores a text-only tensor package with packed U32 weights plus BF16/F32 side tensors." }, "architecture": { "canonical_architecture_id": "kimi-k2-5", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 657.623228416, "main_resident_weight_gb": 655.274418176, "auxiliary_resident_weight_gb": 2.34881024, "fixed_weight_gb": 21.095653376, "routed_expert_weight_gb": 1.6515072, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mlx_u32_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60, model.norm, and lm_head, excluding full input embedding lookup", "auxiliary_scope": "language_model.model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token; no vision_tower or mm_projector tensors are present in this MLX package", "shared_expert_notes": "The config records one shared expert. Shared expert tensors are BF16 and included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the MLX package stores packed U32 switch_mlp expert weights plus BF16 switch_mlp scales and biases, while attention, dense layer 0, shared experts, routers, norms, and lm_head remain BF16/F32. The packed switch_mlp tensors are uniform across the 384 expert dimension, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but constructs expanded key_states and value_states before past_key_values.update. The MLX config records weight quantization only and no KV cache scheme, so Bounds Engine v1 charges expanded two-byte K/V cache streams like the audited base profile." }, "notes": "The served config retains KimiK25ForConditionalGeneration wrapping a DeepseekV3ForCausalLM language model. This MLX artifact's headers contain language_model tensors only; it is modeled as ordinary text decode, not multimodal encoder or prefill throughput." }, "serving": { "weight_format": "mlx_quantized", "weight_bytes_per_param": 0.6407033845077003, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "mlx-lm-4bit-affine-kimi-k2-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored MLX safetensors bytes: packed U32 weights plus BF16 scales, biases, norms, attention tensors, shared experts, embeddings, lm_head, and small F32 router correction biases. Dequantization, activation traffic, Apple MLX scheduling overhead, and compute are outside Bounds Engine v1.", "notes": "The config records 4-bit affine MLX quantization with group_size 32 and mode affine. weight_bytes_per_param records resident stored bytes divided by index total_parameters for catalog display; exact resident, fixed, and routed byte fields drive production bounds." }, "evidence": [ { "label": "MLX Community Kimi K2.5 API metadata", "url": "https://huggingface.co/api/models/mlx-community/Kimi-K2.5", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 351021afd838c866ce1a7374fce51d615773d2a8, the API reports a public non-gated MLX text-generation repo with license other / modified-mit metadata, base_model moonshotai/Kimi-K2.5, base_model:quantized metadata, custom_code, 4-bit and region:us tags, current downloads 225434, and usedStorage 657626220698 bytes. The API does not expose a safetensors parameter summary for this repo." }, { "label": "MLX Community Kimi K2.5 model card", "url": "https://huggingface.co/mlx-community/Kimi-K2.5", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card says this package was converted to MLX format from moonshotai/Kimi-K2.5 using mlx-lm version 0.30.5 with slight modifications, and shows the standard mlx_lm load/generate path." }, { "label": "MLX Community Kimi K2.5 config", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, vocab size 163840, no KV cache scheme, and top-level 4-bit affine MLX quantization with group_size 32." }, { "label": "Moonshot Kimi K2.5 base config comparison", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching checked text-architecture fields between the MLX config and the audited Moonshot base config: model type, architecture, hidden size, layer count, dense-prefix setting, attention heads, KV heads, MLA dimensions, max positions, routed experts, experts per token, shared experts, MoE intermediate size, dense intermediate size, vocab size, and absent KV cache scheme." }, { "label": "MLX Community Kimi K2.5 safetensors index and shard headers", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "The index records total_size 657623228416 bytes and total_parameters 1026408232448 across 182 safetensors shards and 1395 tensors. Range-reading all shard headers matched total_size exactly: U32 507.343011840 GB, BF16 150.280124416 GB, and F32 0.000092160 GB. The package contains no vision_tower or mm_projector tensors. language_model.model.embed_tokens.weight is BF16 [163840, 7168] and contributes 2.348810240 GB resident-only. language_model.lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Packed switch_mlp expert tensors total 634.178764800 GB, exactly 1.651507200 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding and routed switch_mlp tensors is 21.095653376 GB." }, { "label": "MLX Community Kimi K2.5 custom DeepSeek text runtime", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "MLX Community Kimi K2.5 conditional-generation wrapper", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope" ], "notes": "Manual review found the wrapper constructing token embeddings and entering language_model for ordinary text decode. Although the code has multimodal wrapper paths, this MLX package's safetensors headers contain no vision_tower or mm_projector tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the MLX model card, pinned served config, audited Moonshot base config comparison, shipped custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights and missed MLX BF16 scale/bias side tensors, separate lm_head traffic, text-only resident scope, exact routed-expert byte groups, and expanded K/V cache traffic." }, { "id": "moonshotai--kimi-k2-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2.5", "title": "Kimi K2.5 INT4/BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the multimodal Kimi K2.5 compressed-tensors repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-5", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 595.148192736, "main_resident_weight_gb": 591.857094656, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 21.095653376, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_int4_i32_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Because quantization ignores shared_experts, shared expert tensors are BF16 and included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because expert weights are packed INT4 as I32 tensors with BF16 group scales while fixed modules remain mixed BF16/F32. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.562208710190164, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-compressed-tensors-int4-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed INT4 expert payloads, BF16 scales and fixed tensors, and small F32 router correction biases from safetensors headers. Dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The config uses compressed-tensors pack-quantized INT4 weights with group_size 32 and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2.5 model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2.5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 4d01dfe0332d63057c186e0b262165819efb6611, the API reports an image-text-to-text repo with license other, compressed-tensors, custom_code, and safetensors parameters F32: 23040, I32: 1014687129600, BF16: 43902267888, total: 1058589420528. The model card describes Kimi K2.5 as a native multimodal MoE model with 1T total parameters, 32B activated parameters, 61 layers, 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 256K context, MLA attention, and a MoonViT vision encoder. Current downloads were 1,520,063 during audit." }, { "label": "Kimi K2.5 config", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, no next-token-prediction layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, compressed-tensors INT4 pack quantization with group_size 32, kv_cache_scheme null, and MoonViT vision settings." }, { "label": "Kimi K2.5 safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/resolve/4d01dfe0332d63057c186e0b262165819efb6611/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 595148192736 bytes. Range-read shard headers across all 64 shards sum to 595.148192736 GB: BF16 87.804535776 GB, F32 0.00009216 GB, and packed I32 507.3435648 GB. Language tensors excluding the input embedding total 591.857094656 GB; vision_tower contributes 0.833732064 GB; mm_projector contributes 0.108555776 GB; input embedding contributes 2.34881024 GB. Resident-only auxiliary tensors therefore sum to 3.29109808 GB, and ordinary language main resident tensors sum to 591.857094656 GB. Routed expert tensors total 570.76144128 GB, exactly 1.48635792 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095653376 GB." }, { "label": "Kimi K2.5 custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2.5 conditional-generation wrapper", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "auxiliary_resident_scope", "multimodal" ], "notes": "Manual review found the wrapper initializes vision_tower, mm_projector, and language_model. For ordinary text decode without pixel_values, language_model embeddings and decoder drive token generation while vision_tower and mm_projector remain resident but are not swept per generated text token." }, { "label": "Kimi K2.5 deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide documents vLLM, SGLang, and KTransformers serving paths for Kimi K2.5. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, deployment guide, range-read safetensors shard headers, and manual review of the custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official compressed-tensors artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "moonshotai--kimi-k2-6", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2.6", "title": "Kimi K2.6 INT4/BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the multimodal Kimi K2.6 compressed-tensors repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-6", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 595.148192736, "main_resident_weight_gb": 591.857094656, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 21.095653376, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_int4_i32_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Because quantization ignores shared_experts, shared expert tensors are BF16 and included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because expert weights are packed INT4 as I32 tensors with BF16 group scales while fixed modules remain mixed BF16/F32. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.562208710190164, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-compressed-tensors-int4-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed INT4 expert payloads, BF16 scales and fixed tensors, and small F32 router correction biases from safetensors headers. Dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The config uses compressed-tensors pack-quantized INT4 weights with group_size 32 and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2.6 model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2.6", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 7eb5002f6aadc958aed6a9177b7ed26bb94011bb, the API reports an image-text-to-text repo with license other, compressed-tensors, custom_code, and safetensors parameters F32: 23040, I32: 1014687129600, BF16: 43902267888, total: 1058589420528. The model card describes Kimi K2.6 as a native multimodal MoE model with 1T total parameters, 32B activated parameters, 61 layers, 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 256K context, MLA attention, and a MoonViT vision encoder." }, { "label": "Kimi K2.6 config", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/7eb5002f6aadc958aed6a9177b7ed26bb94011bb/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, no next-token-prediction layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, compressed-tensors INT4 pack quantization with group_size 32, kv_cache_scheme null, and MoonViT vision settings." }, { "label": "Kimi K2.6 safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/7eb5002f6aadc958aed6a9177b7ed26bb94011bb/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 595148192736 bytes. Range-read shard headers across all 64 shards sum to 595.148192736 GB: BF16 87.804535776 GB, F32 0.00009216 GB, and packed I32 507.3435648 GB. Language tensors total 594.205904896 GB; vision_tower contributes 0.833732064 GB; mm_projector contributes 0.108555776 GB; input embedding contributes 2.34881024 GB. Resident-only auxiliary tensors therefore sum to 3.29109808 GB, and ordinary language main resident tensors sum to 591.857094656 GB. Routed expert tensors total 570.76144128 GB, exactly 1.48635792 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095653376 GB." }, { "label": "Kimi K2.6 custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/7eb5002f6aadc958aed6a9177b7ed26bb94011bb/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2.6 conditional-generation wrapper", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/7eb5002f6aadc958aed6a9177b7ed26bb94011bb/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "auxiliary_resident_scope", "multimodal" ], "notes": "Manual review found the wrapper initializes vision_tower, mm_projector, and language_model. For ordinary text decode without pixel_values, language_model embeddings and decoder drive token generation while vision_tower and mm_projector remain resident but are not swept per generated text token." }, { "label": "Kimi K2.6 deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/7eb5002f6aadc958aed6a9177b7ed26bb94011bb/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide states Kimi K2.6 shares architecture with Kimi K2.5 and documents vLLM, SGLang, and KTransformers serving paths. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, deployment guide, range-read safetensors shard headers, and manual review of the custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official compressed-tensors artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "moonshotai--kimi-k2-7-code", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2.7-Code", "title": "Kimi K2.7 Code INT4/BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the multimodal Kimi K2.7 Code compressed-tensors repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-7-code", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 595.148192736, "main_resident_weight_gb": 591.857094656, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 21.095653376, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_int4_i32_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Because quantization ignores shared_experts, shared expert tensors are BF16 and included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because expert weights are packed INT4 as I32 tensors with BF16 group scales while fixed modules remain mixed BF16/F32. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.562208710190164, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-compressed-tensors-int4-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed INT4 expert payloads, BF16 scales and fixed tensors, and small F32 router correction biases from safetensors headers. Dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The config uses compressed-tensors pack-quantized INT4 weights with group_size 32 and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2.7 Code model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2.7-Code", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 74797c9c62378b951a1f6fcf5c4631024e9b8bef, the API reports an image-text-to-text repo with license other, compressed-tensors, custom_code, region:us, and safetensors parameters F32: 23040, I32: 1014687129600, BF16: 43902267888, total: 1058589420528. Current downloads were 850031 during audit. The model card describes Kimi K2.7 Code as a native multimodal MoE model with 1T total parameters, 32B activated parameters, 61 layers, 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 256K context, MLA attention, and a MoonViT vision encoder." }, { "label": "Kimi K2.7 Code config", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, no next-token-prediction layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, compressed-tensors INT4 pack quantization with group_size 32, kv_cache_scheme null, and MoonViT vision settings." }, { "label": "Kimi K2.7 Code safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/resolve/74797c9c62378b951a1f6fcf5c4631024e9b8bef/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 595148192736 bytes. Range-read shard headers across all 64 shards sum to 595.148192736 GB: BF16 87.804535776 GB, F32 0.00009216 GB, and packed I32 507.3435648 GB. Language tensors excluding the input embedding total 591.857094656 GB; vision_tower contributes 0.833732064 GB; mm_projector contributes 0.108555776 GB; input embedding contributes 2.34881024 GB. Resident-only auxiliary tensors therefore sum to 3.29109808 GB. Routed expert tensors total 570.76144128 GB, exactly 1.48635792 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095653376 GB." }, { "label": "Kimi K2.7 Code custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2.7 Code conditional-generation wrapper", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "auxiliary_resident_scope", "multimodal" ], "notes": "Manual review found the wrapper initializes vision_tower, mm_projector, and language_model. For ordinary text decode without pixel_values, language_model embeddings and decoder drive token generation while vision_tower and mm_projector remain resident but are not swept per generated text token." }, { "label": "Kimi K2.7 Code deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide states Kimi K2.7 Code has the same architecture as Kimi K2.5/Kimi K2.6 and documents vLLM, SGLang, and KTransformers serving paths. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the served config, model card, deployment guide, range-read safetensors shard headers, and manual review of the shipped custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official compressed-tensors artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "moonshotai--kimi-k2-instruct-0905", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2-Instruct-0905", "title": "Kimi K2 Instruct 0905 FP8", "summary": "Audited memory-side ordinary text-decode bounds profile for the official Kimi K2 Instruct 0905 block-FP8 repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-instruct-0905", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1029.173265504, "main_resident_weight_gb": 1026.824455264, "auxiliary_resident_weight_gb": 2.34881024, "fixed_weight_gb": 11.890705504, "routed_expert_weight_gb": 2.64305664, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.layers 0-60, excluding full input embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Shared expert tensors are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the block-FP8 package mixes F8_E4M3 tensors with BF16 embeddings/norms and F32 scale-inverse tensors. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "DeepseekV3ForCausalLM text generation profile for the official Kimi K2 Instruct 0905 checkpoint. This profile models ordinary text decode; tool parsing, prefill scheduling, and optimized runtime-specific cache compression are outside this v1 bound." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0026328369262476, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-block-fp8-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weight payloads, F32 scale-inverse tensors, BF16 embeddings/norms/output head, and expanded BF16 K/V cache bytes from the custom code path. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config uses FP8 e4m3 block quantization with 128x128 weight blocks and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2 Instruct 0905 model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Instruct-0905", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit ac6c49f04883bd0a0598b790693a72061c676629, the API reports a text-generation repo with license other, custom_code, fp8 tag, and safetensors parameters F32: 62518232, BF16: 2514975360, F8_E4M3: 1023893241856, total: 1026470735448. The model card describes Kimi K2-Instruct-0905 as a MoE language model with 1T total parameters, 32B activated parameters, 61 layers including 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 256K context, MLA attention, and block-FP8 checkpoints." }, { "label": "Kimi K2 Instruct 0905 config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/ac6c49f04883bd0a0598b790693a72061c676629/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records model_type kimi_k2, DeepseekV3ForCausalLM, torch_dtype bfloat16, 61 hidden layers, no next-token-prediction layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 block quantization, and kv_cache_scheme null." }, { "label": "Kimi K2 Instruct 0905 safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/ac6c49f04883bd0a0598b790693a72061c676629/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 1029173265504 bytes. Range-read shard headers across all 62 shards sum to 1029.173265504 GB: F8_E4M3 1023.893241856 GB, BF16 5.02995072 GB, and F32 0.250072928 GB. The total parameter count from headers is 1026470735448, matching the API safetensors total. model.embed_tokens.weight contributes 2.34881024 GB and is resident-only for ordinary text decode. lm_head.weight is separate and remains swept. Routed expert tensors total 1014.93374976 GB, exactly 2.64305664 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 11.890705504 GB." }, { "label": "Kimi K2 Instruct 0905 custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/ac6c49f04883bd0a0598b790693a72061c676629/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2 Instruct 0905 deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/ac6c49f04883bd0a0598b790693a72061c676629/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide states the smallest Kimi-K2 FP8 deployment unit with 256K sequence length on mainstream H200 platforms is 16 GPUs and documents vLLM, SGLang, KTransformers, and TensorRT-LLM serving paths. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, deployment guide, range-read safetensors shard headers, and manual review of the custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official block-FP8 artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "moonshotai--kimi-k2-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2-Instruct", "title": "Kimi K2 Instruct FP8", "summary": "Audited memory-side ordinary text-decode bounds profile for the official original Kimi K2 Instruct block-FP8 repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-instruct", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1029.17325672, "main_resident_weight_gb": 1026.82444648, "auxiliary_resident_weight_gb": 2.34881024, "fixed_weight_gb": 11.89069672, "routed_expert_weight_gb": 2.64305664, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.layers 0-60, excluding full input embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Shared expert tensors are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the block-FP8 package mixes F8_E4M3 tensors with BF16 embeddings/norms and F32 scale-inverse tensors. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "DeepseekV3ForCausalLM text generation profile for the original official Kimi K2 Instruct checkpoint. This profile models ordinary text decode; tool parsing, prefill scheduling, and optimized runtime-specific cache compression are outside this v1 bound." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0026328326587741, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-block-fp8-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weight payloads, F32 scale-inverse tensors, BF16 embeddings/norms/output head, and expanded BF16 K/V cache bytes from the custom code path. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config uses FP8 e4m3 block quantization with 128x128 weight blocks and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2 Instruct model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit fd1984e2b7a3350dbf7305fe73a4ede25c14de50, the API reports a text-generation repo with license other, custom_code, fp8 tag, region:us, and safetensors parameters F32: 62518232, BF16: 2514970968, F8_E4M3: 1023893241856, total: 1026470731056. Current downloads were 379,918 during audit. The model card describes Kimi K2 as a MoE language model with 1T total parameters, 32B activated parameters, 61 layers including 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 128K context, MLA attention, and block-FP8 checkpoints." }, { "label": "Kimi K2 Instruct config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/raw/fd1984e2b7a3350dbf7305fe73a4ede25c14de50/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records model_type kimi_k2, DeepseekV3ForCausalLM, torch_dtype bfloat16, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 131072, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, FP8 e4m3 block quantization, and rope scaling factor 32." }, { "label": "Kimi K2 Instruct safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/resolve/fd1984e2b7a3350dbf7305fe73a4ede25c14de50/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 1029173256720 bytes. Range-read shard headers across all 61 shards sum to 1029.173256720 GB: F8_E4M3 1023.893241856 GB, BF16 5.029941936 GB, and F32 0.250072928 GB. The total parameter count from headers is 1026470731056, matching the API safetensors total. model.embed_tokens.weight contributes 2.348810240 GB and is resident-only for ordinary text decode. lm_head.weight is separate and remains swept. Routed expert tensors total 1014.933749760 GB, exactly 2.643056640 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 11.890696720 GB." }, { "label": "Kimi K2 Instruct custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/raw/fd1984e2b7a3350dbf7305fe73a4ede25c14de50/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2 Instruct deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/raw/fd1984e2b7a3350dbf7305fe73a4ede25c14de50/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide states the smallest Kimi-K2 FP8 deployment unit with 128K sequence length on mainstream H200/H20 platforms is 16 GPUs and documents vLLM, SGLang, KTransformers, and TensorRT-LLM serving paths. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from the served config, model card/API metadata, deployment guide, range-read safetensors shard headers, and manual review of the custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official original block-FP8 artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "moonshotai--kimi-k2-thinking", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-K2-Thinking", "title": "Kimi K2 Thinking INT4/BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the official Kimi K2 Thinking compressed-tensors repo.", "model_family": "kimi-k2-moe", "architecture": { "canonical_architecture_id": "kimi-k2-thinking", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 594.205920512, "main_resident_weight_gb": 591.857110272, "auxiliary_resident_weight_gb": 2.34881024, "fixed_weight_gb": 21.095668992, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_int4_i32_bf16_f32", "traffic_scope": "ordinary text decode through model.layers 0-60, excluding full input embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Because quantization ignores shared_experts, shared expert tensors are BF16 and included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the compressed-tensors package stores packed INT4 expert payloads as I32 tensors with BF16 group scales while fixed modules remain mixed BF16/F32. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but expands key_states and value_states before past_key_values.update, so Bounds Engine v1 charges full-context BF16 K and V cache streams rather than a latent MLA cache." }, "notes": "DeepseekV3ForCausalLM text generation profile for the official Kimi K2 Thinking checkpoint. This profile models ordinary text decode; reasoning parser behavior, tool parsing, prefill scheduling, and optimized runtime-specific cache compression are outside this v1 bound." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.561568521462566, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-custom-code-compressed-tensors-int4-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed INT4 expert payloads, BF16 scales and fixed tensors, small F32 router correction biases, and expanded BF16 K/V cache bytes from the custom code path. Dequantization, activation traffic, reasoning/tool parser overhead, and compute throughput are outside Bounds Engine v1.", "notes": "The config uses compressed-tensors pack-quantized INT4 weights with group_size 32 and kv_cache_scheme null. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi K2 Thinking model card and API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Thinking", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55, the API reports a public non-gated text-generation repo with license other, compressed-tensors, custom_code, region:us, downloads 139650, usedStorage 594283168582 bytes, and safetensors parameters F32: 23040, I32: 1014687129600, BF16: 43431131776, total: 1058118284416. The model card describes Kimi K2 Thinking as a native INT4 MoE thinking model with 1T total parameters, 32B activated parameters, 61 layers including 1 dense layer, hidden size 7168, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocab, 256K context, MLA attention, and vLLM/SGLang/KTransformers serving." }, { "label": "Kimi K2 Thinking config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records model_type kimi_k2, DeepseekV3ForCausalLM, torch_dtype bfloat16, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, compressed-tensors INT4 pack quantization with group_size 32, and kv_cache_scheme null." }, { "label": "Kimi K2 Thinking safetensors index and shard headers", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/resolve/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 594205920512 bytes. Range-read shard headers across all 62 shards sum to 594.205920512 GB across 208276 tensors: BF16 86.862263552 GB, F32 0.000092160 GB, and packed I32 507.343564800 GB. The total parameter count from headers matches the API safetensors total. model.embed_tokens.weight contributes 2.348810240 GB and is resident-only for ordinary text decode. lm_head.weight is separate and remains swept. Routed expert tensors total 570.761441280 GB, exactly 1.486357920 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095668992 GB." }, { "label": "Kimi K2 Thinking custom DeepSeek text runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The file is byte-identical to the audited Kimi K2 Instruct 0905 modeling_deepseek.py for the attention/cache path, and the ordinary text decoder stack uses 61 layers." }, { "label": "Kimi K2 Thinking deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "runtime_scope" ], "notes": "The deployment guide states the smallest Kimi K2 Thinking INT4 deployment unit with 256K sequence length on mainstream H200 platforms is 8 GPUs with tensor parallelism, and documents vLLM, SGLang, and KTransformers serving paths. This profile deliberately models the audited custom-code expanded K/V cache path until a runtime-specific compressed-cache implementation is separately audited." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card text, pinned served config, deployment guide, resolved safetensors index, direct range-read safetensors shard headers, and manual review of the custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the official compressed-tensors artifact, not a claim about optimized runtime-specific MLA cache compression or reasoning/tool parser overhead." }, { "id": "moonshotai--kimi-vl-a3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-VL-A3B-Instruct", "title": "Kimi VL A3B Instruct BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 Kimi-VL A3B Instruct multimodal MoE repo.", "model_family": "kimi-vl-a3b-moe", "base_model_proof": { "base_model": "moonshotai/Moonlight-16B-A3B", "relation": "finetune", "source": "Hugging Face base_model metadata plus direct text-config comparison", "config_compatible": false, "notes": "The repo declares Moonlight-16B-A3B as its fine-tuning base. Manual comparison found matching text decoder geometry except Kimi-VL extends max_position_embeddings from 8192 to 131072 and changes rope_theta from 50000 to 800000. Kimi-VL also adds a MoonViT vision tower and multimodal projector, so this profile audits the served wrapper directly." }, "architecture": { "canonical_architecture_id": "kimi-vl-a3b-instruct", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.815315552, "main_resident_weight_gb": 31.249135232, "auxiliary_resident_weight_gb": 1.56618032, "fixed_weight_gb": 2.459432576, "routed_expert_weight_gb": 0.449839104, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through language_model.model.layers.*, language_model.model.norm.weight, and language_model.lm_head.weight, excluding the input embedding matrix and resident multimodal tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower tensors, and multi_modal_projector tensors are resident for the multimodal package but are not swept as full matrices for each generated text token", "shared_expert_notes": "The text config records two shared experts. Shared expert tensors are included in fixed_weight_gb because the runtime adds shared_experts every MoE layer.", "notes": "Range-read safetensors headers record 7 shards and 5679 tensors totaling 32.815315552 GB, all BF16. Routed expert tensors under language_model.model.layers.1-26.mlp.experts.* are byte-uniform across all 64 expert indexes at exactly 0.449839104 GB per expert index. Non-routed ordinary text traffic totals 2.459432576 GB. Resident-only input embedding, vision tower, and multimodal projector tensors total 1.566180320 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeekV3 attention computes compressed_kv internally but constructs expanded key_states and value_states before past_key_value.update. Bounds Engine v1 therefore charges full-context BF16 K and V cache streams rather than a latent compressed cache." }, "notes": "KimiVLForConditionalGeneration combines a MoonViT vision tower, multimodal projector, and DeepseekV3ForCausalLM language model. This profile models generated text-token decode after any image/video prefill, not vision encoder, projector, media tiling, or prefill throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "kimi-vl-transformers-bf16-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, multimodal projector prefill, image/video processing, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16 and all safetensors payload bytes are BF16. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi VL A3B Instruct API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-VL-A3B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At repo SHA 398eede0903cd983a2bfa0cc634e9ac1d843f375, the API reports a public non-gated transformers image-text-to-text repo with MIT license metadata, custom code, base_model moonshotai/Moonlight-16B-A3B with base_model:finetune metadata, region:us, 332831 downloads, usedStorage 32822309117 bytes, and safetensors parameters BF16 16407657776." }, { "label": "Kimi VL A3B Instruct model card", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/README.md", "source_type": "model_card", "supports": [ "model_family", "pipeline", "license", "multimodal_scope", "serving" ], "notes": "The card identifies Kimi-VL as an MoE vision-language model with a native-resolution MoonViT visual encoder and MLP projector. It states 16B total parameters, 3B activated parameters, 128K context length, advanced multimodal and long-context capability, and recommends Kimi-VL-A3B-Instruct for general multimodal perception, OCR, long video, long document, video perception, and agent use." }, { "label": "Kimi VL A3B Instruct served config", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "vision_projector" ], "notes": "The served config records KimiVLForConditionalGeneration, model_type kimi_vl, bfloat16, text_config DeepseekV3 with 27 layers, first_k_dense_replace 1, hidden size 2048, intermediate size 11264, MoE intermediate size 1408, 16 attention heads, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, untied embeddings, vocab size 163840, and 131072 max position embeddings. The vision_config records MoonViT with 27 layers, hidden size 1152, 16 attention heads, patch size 14, and merge kernel size 2x2." }, { "label": "Moonlight 16B A3B base config comparison", "url": "https://huggingface.co/moonshotai/Moonlight-16B-A3B/raw/476b36a473d4467f94469414bef6cee75c9c8172/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found Kimi-VL text_config and the Moonlight base config have the same text geometry for hidden size, layer count, MoE expert counts, MLA dimensions, attention heads, vocabulary size, and untied embeddings. They differ in max_position_embeddings, which Kimi-VL extends from 8192 to 131072, and rope_theta, which Kimi-VL changes from 50000 to 800000." }, { "label": "Kimi VL custom runtime", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/modeling_kimi_vl.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "shared_experts", "multimodal_scope", "ordinary_text_decode_scope" ], "notes": "Manual review found KimiVLForConditionalGeneration builds vision_tower, multi_modal_projector, and language_model = DeepseekV3ForCausalLM. Forward embeds text tokens, runs the vision tower and projector when pixel_values are present, merges image features into input embeddings, and delegates generation to the language model. DeepseekV3Attention computes compressed_kv but expands key_states to 16 heads x 192 dimensions and value_states to 16 heads x 128 dimensions before past_key_value.update in the ordinary eager path. DeepseekV3MoE routes top-k experts and adds shared_experts every MoE layer." }, { "label": "Kimi VL A3B Instruct safetensors headers", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/resolve/398eede0903cd983a2bfa0cc634e9ac1d843f375/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "All seven safetensors shard headers were range-read directly. Stored tensor payloads sum to 32.815315552 GB across 5679 tensors, all BF16, matching index total_size. The language input embedding is 0.671088640 GB, vision_tower tensors total 0.833732064 GB, and multi_modal_projector tensors total 0.061359616 GB, giving 1.566180320 GB resident-only auxiliary payload. Language tensors excluding input embedding total 31.249135232 GB. Routed expert tensors total 28.789702656 GB, exactly 0.449839104 GB per expert index across all 64 experts and 26 MoE layers. Fixed ordinary-decode traffic, including dense layer 0 MLP, attention, routers, shared experts, norms, and lm_head, totals 2.459432576 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, Moonlight base config comparison, custom runtime code, safetensors index, and direct range-read safetensors shard header grouping." }, "notes": "This profile supersedes the generated metadata estimate. It is an ordinary text-token decode profile after multimodal prefill and does not claim image/video encoder throughput or runtime-specific latent MLA cache compression." }, { "id": "moonshotai--kimi-vl-a3b-thinking", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "moonshotai/Kimi-VL-A3B-Thinking", "title": "Kimi VL A3B Thinking BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 Kimi-VL A3B Thinking multimodal MoE repo.", "model_family": "kimi-vl-a3b-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-VL-A3B-Instruct", "relation": "finetune", "source": "Hugging Face base_model metadata, byte-identical served config comparison against the audited Instruct profile, runtime diff review, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo declares moonshotai/Kimi-VL-A3B-Instruct as its fine-tuning base. Manual comparison found config.json and configuration_kimi_vl.py are byte-identical to the audited Instruct repo. modeling_kimi_vl.py differs only by replacing a fixed _supports_sdpa = False class attribute with a property that delegates to language_model._supports_sdpa; the attention/cache path, MoE routing, vision tower, projector, and ordinary text generation dataflow are unchanged for Bounds Engine v1." }, "architecture": { "canonical_architecture_id": "kimi-vl-a3b-thinking", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.815315552, "main_resident_weight_gb": 31.249135232, "auxiliary_resident_weight_gb": 1.56618032, "fixed_weight_gb": 2.459432576, "routed_expert_weight_gb": 0.449839104, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through language_model.model.layers.*, language_model.model.norm.weight, and language_model.lm_head.weight, excluding the input embedding matrix and resident multimodal tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower tensors, and multi_modal_projector tensors are resident for the multimodal package but are not swept as full matrices for each generated text token", "shared_expert_notes": "The text config records two shared experts. Shared expert tensors are included in fixed_weight_gb because the runtime adds shared_experts every MoE layer.", "notes": "Range-read safetensors headers record 7 shards and 5679 tensors totaling 32.815315552 GB, all BF16. Routed expert tensors under language_model.model.layers.1-26.mlp.experts.* are byte-uniform across all 64 expert indexes at exactly 0.449839104 GB per expert index. Non-routed ordinary text traffic totals 2.459432576 GB. Resident-only input embedding, vision tower, and multimodal projector tensors total 1.566180320 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeekV3 attention computes compressed_kv internally but constructs expanded key_states and value_states before past_key_value.update. Bounds Engine v1 therefore charges full-context BF16 K and V cache streams rather than a latent compressed cache." }, "notes": "KimiVLForConditionalGeneration combines a MoonViT vision tower, multimodal projector, and DeepseekV3ForCausalLM language model. This profile models generated text-token decode after any image/video prefill, not vision encoder, projector, media tiling, or prefill throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "kimi-vl-transformers-bf16-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, multimodal projector prefill, image/video processing, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16 and all safetensors payload bytes are BF16. The audited custom code path stores expanded BF16 K/V cache tensors." }, "evidence": [ { "label": "Kimi VL A3B Thinking API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-VL-A3B-Thinking", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At repo SHA 7d99e220af610d8624fcba22b2c076c7ed528f14, the API reports a public non-gated transformers image-text-to-text repo with MIT license metadata, custom code, base_model moonshotai/Kimi-VL-A3B-Instruct with base_model:finetune metadata, region:us, 145860 downloads, usedStorage 32820262484 bytes, and safetensors parameters BF16 16407657776." }, { "label": "Kimi VL A3B Thinking model card", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/raw/7d99e220af610d8624fcba22b2c076c7ed528f14/README.md", "source_type": "model_card", "supports": [ "model_family", "pipeline", "license", "multimodal_scope", "serving" ], "notes": "The card identifies Kimi-VL as a 16B total / 3B active MoE vision-language model with native-resolution MoonViT, MLP projector, and 128K context. It describes Kimi-VL-Thinking as a long chain-of-thought SFT/RL variant for advanced text and multimodal reasoning while retaining the compact A3B activated language decoder footprint." }, { "label": "Kimi VL A3B Thinking served config", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/raw/7d99e220af610d8624fcba22b2c076c7ed528f14/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "vision_projector" ], "notes": "The served config is byte-identical to the audited Kimi-VL-A3B-Instruct config and records KimiVLForConditionalGeneration, model_type kimi_vl, bfloat16, text_config DeepseekV3 with 27 layers, first_k_dense_replace 1, hidden size 2048, intermediate size 11264, MoE intermediate size 1408, 16 attention heads, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, untied embeddings, vocab size 163840, and 131072 max position embeddings. The vision_config records MoonViT with 27 layers, hidden size 1152, 16 attention heads, patch size 14, and merge kernel size 2x2." }, { "label": "Kimi VL A3B Instruct audited profile and config comparison", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter", "multimodal_scope" ], "notes": "Manual comparison found Thinking config.json and configuration_kimi_vl.py are byte-identical to the audited Instruct repo. The already audited Instruct profile established the Kimi-VL text decoder geometry, MoonViT vision tower, multimodal projector, expanded BF16 K/V cache path, and ordinary text-decode scope." }, { "label": "Kimi VL A3B Thinking custom runtime diff review", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/raw/7d99e220af610d8624fcba22b2c076c7ed528f14/modeling_kimi_vl.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "shared_experts", "multimodal_scope", "ordinary_text_decode_scope" ], "notes": "Manual diff against the audited Instruct runtime found only an _supports_sdpa exposure change in KimiVLPreTrainedModel. The ordinary generation dataflow remains the same: optional vision_tower and multi_modal_projector execution for media inputs, then language_model = DeepseekV3ForCausalLM. The DeepseekV3Attention cache path expands keys to 16 heads x 192 dimensions and values to 16 heads x 128 dimensions before past_key_value.update, and DeepseekV3MoE routes top-k experts plus shared_experts every MoE layer." }, { "label": "Kimi VL A3B Thinking safetensors headers", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/resolve/7d99e220af610d8624fcba22b2c076c7ed528f14/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "All seven safetensors shard headers were range-read directly. Stored tensor payloads sum to 32.815315552 GB across 5679 tensors, all BF16, matching index total_size. The language input embedding is 0.671088640 GB, vision_tower tensors total 0.833732064 GB, and multi_modal_projector tensors total 0.061359616 GB, giving 1.566180320 GB resident-only auxiliary payload. Language tensors excluding input embedding total 31.249135232 GB. Routed expert tensors total 28.789702656 GB, exactly 0.449839104 GB per expert index across all 64 experts and 26 MoE layers. Fixed ordinary-decode traffic, including dense layer 0 MLP, attention, routers, shared experts, norms, and lm_head, totals 2.459432576 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, byte-identical configuration comparison against the audited Instruct repo, custom runtime diff review, safetensors index, and direct range-read safetensors shard header grouping." }, "notes": "This profile supersedes the generated metadata estimate. It is an ordinary text-token decode profile after multimodal prefill and does not claim image/video encoder throughput or runtime-specific latent MLA cache compression." }, { "id": "mratsim--glm-4-32b-0414-w4a16-gptq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "mratsim/GLM-4-32B-0414.w4a16-gptq", "title": "mratsim GLM 4 32B 0414 GPTQ W4A16", "summary": "Audited memory-side text-decode bounds profile for the mratsim asymmetric GPTQ W4A16 package of GLM-4-32B-0414.", "model_family": "glm4-dense-gptq", "base_model_proof": { "base_model": "zai-org/GLM-4-32B-0414", "relation": "quantized", "source": "Hugging Face model card/API metadata, served GPTQ config, base config comparison, quantization recipe, and direct safetensors header review", "config_compatible": true, "notes": "The API and model card identify this repo as a quantized derivative of zai-org/GLM-4-32B-0414. Manual comparison found the served GPTQ config preserves the base GLM4 architecture fields used by this profile: Glm4ForCausalLM, hidden size, MLP size, layer count, attention head count, KV head count, head dimension, context length, BF16 dtype, untied embeddings, vocabulary size, RoPE theta, and partial rotary factor. The repo adds compressed-tensors GPTQ W4A16 metadata." }, "architecture": { "canonical_architecture_id": "glm-4-32b-0414", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.566081536, "swept_params_b": 31.634946048, "auxiliary_resident_params_b": 0.931135488, "resident_weight_gb": 19.678365408, "swept_weight_gb": 17.816094432, "auxiliary_resident_weight_gb": 1.862270976, "resident_parameter_scope": "logical GLM-4-32B-0414 parameters from the served/base config plus direct GPTQ safetensors stored-byte totals", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 weight_packed and weight_zero_point tensors, BF16 weight_scale and unquantized tensors, and I64 weight_shape tensors. Logical parameter counts use the base GLM-4-32B-0414 geometry; storage metadata tensors are byte traffic, not separate logical model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 61, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 61 decoder layers, 2 KV heads, 128-dimensional key/value heads, 32768 max positions, and no sliding-window or latent-cache setting. The Transformers GLM4 attention path updates standard key/value cache entries, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "Dense Glm4ForCausalLM profile for ordinary cached text decode. The model card documents a serving example with YaRN rope scaling to 130000 tokens, but the committed config's max_position_embeddings is 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6042595387549667, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-gptq-glm4-w4a16-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ safetensors bytes: packed I32 weights/zero points, BF16 scales, BF16 unquantized tensors, and I64 shape tensors. Dequantization, activation traffic, vLLM kernel behavior, scheduler behavior, prefix caching, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized int4 weights with group size 128, asymmetric quantization, and lm_head ignored. The README describes the artifact as 4-bit weight-only W4A16 GPTQ and recommends vLLM. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "mratsim GLM 4 32B 0414 GPTQ API metadata", "url": "https://huggingface.co/api/models/mratsim/GLM-4-32B-0414.w4a16-gptq", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 7649a2140aa2dd535d23c8794be7662f23a7aee8, the API records a public MIT text-generation safetensors repo with glm4, gptq, vllm, llmcompressor, text-generation-inference, compressed-tensors, region:us, and base_model zai-org/GLM-4-32B-0414 tags. Current downloads were 74733 during audit, while the local catalog snapshot retains 148059. The API safetensors block reports I32, BF16, and I64 storage-accounting entries; this profile uses direct header byte spans for traffic." }, { "label": "mratsim GLM 4 32B 0414 GPTQ model card", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/raw/7649a2140aa2dd535d23c8794be7662f23a7aee8/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format", "serving", "quantization_recipe" ], "notes": "The README states this repo contains GLM-4-32B-0414 quantized with asymmetric GPTQ to 4-bit W4A16 for consumer hardware. It records calibration on 2048 samples with max sequence length 4096 from mit-han-lab/pile-val-backup and provides a vLLM serving example." }, { "label": "mratsim GLM 4 32B 0414 GPTQ config", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/raw/7649a2140aa2dd535d23c8794be7662f23a7aee8/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "embedding_layout", "serving", "kv_adapter" ], "notes": "The served config records Glm4ForCausalLM, model_type glm4, torch_dtype bfloat16, hidden size 6144, intermediate size 23040, 61 layers, 48 attention heads, 2 KV heads, head_dim 128, 32768 max position embeddings, tie_word_embeddings false, vocab size 151552, partial_rotary_factor 0.5, rope_theta 10000, and compressed-tensors pack-quantized 4-bit int weights with group size 128, asymmetric quantization, and lm_head in the ignore list." }, { "label": "mratsim GLM 4 32B 0414 GPTQ recipe", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/raw/7649a2140aa2dd535d23c8794be7662f23a7aee8/recipe.yaml", "source_type": "config", "supports": [ "serving", "weight_format", "quantization_recipe" ], "notes": "The recipe applies GPTQModifier to Linear targets with 4-bit integer weights, group size 128, asymmetric minmax observer, group strategy, dampening_frac 0.005, and ignore [lm_head]." }, { "label": "GLM 4 32B 0414 base API and config", "url": "https://huggingface.co/zai-org/GLM-4-32B-0414/raw/077b5c2f5c43bd3239fd605a0600229e8facbd4a/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_params_b", "embedding_layout", "kv_adapter" ], "notes": "The base API records MIT licensing, BF16 safetensors total 32566081536 logical parameters, and public Transformers text-generation metadata. Manual config comparison found matching model type, dtype, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, context length, untied embeddings, vocabulary size, RoPE theta, partial rotary factor, attention bias, use_cache, and RMS norm epsilon between the base config and GPTQ config." }, { "label": "mratsim GLM 4 32B 0414 GPTQ safetensors index and shard headers", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/resolve/7649a2140aa2dd535d23c8794be7662f23a7aee8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The index maps 1711 tensors across four safetensors shards and records total_size 19678365408 bytes. Direct range-read shard headers match exactly: tensor payloads total 19.678365408 GB, split into I32 15.471083520 GB, BF16 4.207276032 GB, and I64 0.000005856 GB. Stored suffix totals are weight_packed 15.351152640 GB, weight_zero_point 0.119930880 GB, weight_scale 0.479723520 GB, weight_shape 0.000005856 GB, and unquantized BF16 weight tensors 3.727552512 GB. model.embed_tokens.weight is BF16 with shape [151552, 6144] and contributes 1.862270976 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.layers.* plus model.norm.weight plus lm_head.weight sum to 17.816094432 GB swept traffic. The four linked shard files total 19.678565656 GB, leaving 0.000200248 GB of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers GLM4 attention implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.51.3/src/transformers/models/glm4/modeling_glm4.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual implementation review found Glm4Attention defines k_proj and v_proj with num_key_value_heads * head_dim, applies RoPE to key states, and calls past_key_value.update(key_states, value_states, layer_idx, cache_kwargs). This supports charging standard full-context BF16 key/value cache streams for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, README, served GPTQ config, recipe.yaml, base API/config comparison, safetensors index, direct range-read safetensors shard headers, linked-object HEAD checks, and the Transformers GLM4 runtime implementation." }, "notes": "This profile supersedes the scraped flat 4-bit estimate by using exact stored GPTQ bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "mungert--hunyuan-mt-7b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Mungert/Hunyuan-MT-7B-GGUF", "title": "Mungert Hunyuan-MT-7B GGUF BF16/Q8_0", "summary": "Audited memory-side text-decode bounds profile for the API-selected Hunyuan-MT-7B BF16/Q8_0 GGUF artifact.", "model_family": "hunyuan-mt-7b-dense-gguf", "base_model_proof": { "base_model": "tencent/Hunyuan-MT-7B", "relation": "quantized", "source": "Mungert model card, Tencent base config, live HF GGUF metadata, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": false, "notes": "The Mungert card bundles the original Tencent Hunyuan-MT-7B card and identifies these files as GGUF conversions produced with llama.cpp. The selected GGUF header matches the Tencent base on dense decoder geometry: 32 layers, hidden size 4096, feed-forward size 14336, 32 attention heads, 8 KV heads, 128 key/value head dimensions, vocabulary size 128256, tied embeddings, and qk norms. The checked context field differs: Tencent config max_position_embeddings is 32768, while the selected GGUF header and live API record 262144 context, so this profile treats the selected GGUF header as the source of truth for context." }, "architecture": { "canonical_architecture_id": "hunyuan-mt-7b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.504932864, "swept_params_b": 7.504932864, "auxiliary_resident_params_b": 0, "resident_weight_gb": 10.865575104, "swept_weight_gb": 10.858053632, "auxiliary_resident_weight_gb": 0.007521472, "resident_parameter_scope": "selected GGUF linked file size and GGUF header tensor parameters for Hunyuan-MT-7B-bf16_q8_0.gguf", "swept_parameter_scope": "ordinary text decode charges all selected GGUF tensor spans because the tied token_embd.weight tensor is also the output projection and there is no separate output.weight tensor", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file alignment are resident in the selected artifact but are not model tensor traffic", "notes": "The live HF API gguf.totalFileSize matches both Hunyuan-MT-7B-bf16_q8_0.gguf and Hunyuan-MT-7B-f16_q8_0.gguf. The two files have identical memory-side byte geometry: 354 tensors, 7.504932864B logical tensor elements, 10.858053632 GB tensor spans, 0.007521472 GB metadata/header/alignment overhead, and 10.865575104 GB linked file size. This profile targets the first API-size-matching BF16/Q8_0 artifact and notes that the F16/Q8_0 sibling has the same memory-side bound." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records hunyuan-dense.block_count 32, attention.head_count 32, attention.head_count_kv 8, key_length 128, value_length 128, and context_length 262144." }, "notes": "Hunyuan-MT-7B is a dense decoder-only causal LM used for translation prompts. This profile models ordinary text decode for the selected GGUF artifact; other quantized siblings such as Q8_0, Q4_K_M, and IQ4_NL have different resident and swept bytes." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 1.447791112978569, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-q8-0-hunyuan-dense-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, kernels, dequantization, sampling, scheduler behavior, and prompt-template processing are outside Bounds Engine v1.", "notes": "The selected artifact stores BF16 high-precision tensors, Q8_0 quantized tensors, and F32 norms. The GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "evidence": [ { "label": "Mungert Hunyuan-MT-7B GGUF API metadata", "url": "https://huggingface.co/api/models/Mungert/Hunyuan-MT-7B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "pipeline", "selected_artifact", "total_params_b", "max_context_tokens", "license" ], "notes": "At commit 61e98ae605cc4fe9581fd3ff1052a271843e4d64, the live API records a public non-gated translation GGUF repo with transformers, gguf, translation, endpoints_compatible, region:us, conversational, and multilingual tags. Current downloads are 170812. The API GGUF block records architecture hunyuan-dense, context_length 262144, total 7504932864, and totalFileSize 10865575104." }, { "label": "Mungert Hunyuan-MT-7B GGUF model card", "url": "https://huggingface.co/Mungert/Hunyuan-MT-7B-GGUF/raw/61e98ae605cc4fe9581fd3ff1052a271843e4d64/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "pipeline", "language", "serving" ], "notes": "The card says the GGUF files were generated with llama.cpp commit c8dedc9999eccf7821a9fe5b29f10e8d075e2217 and bundles the original Hunyuan-MT model card. The original card identifies Hunyuan-MT-7B as the Tencent 7B translation model, provides translation prompt templates, and links the tencent/Hunyuan-MT-7B base model." }, { "label": "Tencent Hunyuan-MT-7B base config", "url": "https://huggingface.co/tencent/Hunyuan-MT-7B/raw/9305c78383f0bcc94358e08667ee2c76107877e3/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "embedding_layout", "serving" ], "notes": "The pinned base config records HunYuanDenseV1ForCausalLM, hunyuan_v1_dense, bfloat16 source dtype, 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, qk norm enabled, tied word embeddings, vocab size 128256, use_cache true, and max_position_embeddings 32768." }, { "label": "Mungert Hunyuan-MT-7B GGUF linked-object HEAD checks", "url": "https://huggingface.co/Mungert/Hunyuan-MT-7B-GGUF/tree/61e98ae605cc4fe9581fd3ff1052a271843e4d64", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Hunyuan-MT-7B-bf16_q8_0.gguf and Hunyuan-MT-7B-f16_q8_0.gguf are each 10865575104 bytes, exactly matching API gguf.totalFileSize. Other siblings differ: Hunyuan-MT-7B-bf16.gguf is 15017936064 bytes, q8_0 is 7982318784 bytes, q6_k_m is 6292014560 bytes, q5_k_m is 5516592608 bytes, and q4_k_m is 4702111200 bytes." }, { "label": "Mungert Hunyuan-MT-7B BF16/Q8_0 GGUF range-read tensor index", "url": "https://huggingface.co/Mungert/Hunyuan-MT-7B-GGUF/resolve/61e98ae605cc4fe9581fd3ff1052a271843e4d64/Hunyuan-MT-7B-bf16_q8_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "embedding_layout" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 34 metadata entries and 354 tensors. The linked file is 10.865575104 GB. Tensor spans sum to 10.858053632 GB across 7.504932864B logical elements, while metadata/header/alignment padding accounts for 0.007521472 GB. Tensor spans split into BF16 6.150946816 GB, Q8_0 4.706009088 GB, and F32 0.001097728 GB. token_embd.weight is 1.050673152 GB and is swept because the base config ties embeddings and the GGUF header has no separate output.weight tensor. blk.* tensors total 9.807364096 GB and output_norm.weight is 0.000016384 GB. The header records hunyuan-dense.block_count 32, context_length 262144, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, attention.key_length 128, attention.value_length 128, and rope.scaling.original_context_length 262144." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the pinned Mungert model card, Tencent base config/API metadata, linked-object HEAD checks, and direct GGUF header/tensor-index range reads of the API-size-matching BF16/Q8_0 and F16/Q8_0 artifacts." }, "notes": "Use this profile for the API-selected Hunyuan-MT-7B BF16/Q8_0 GGUF memory-side bound. Do not silently substitute the smaller Q8_0, Q4_K_M, or IQ4_NL siblings; those require separate selected-artifact profiles." }, { "id": "nm-testing--qwen1-5-moe-a2-7b-chat-quantized-w4a16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16", "title": "NM Testing Qwen1.5 MoE A2.7B Chat W4A16", "summary": "Audited memory-side text-decode bounds profile for the nm-testing compressed-tensors W4A16 package of Qwen1.5 MoE A2.7B Chat.", "model_family": "qwen2-moe", "base_model_proof": { "base_model": "Qwen/Qwen1.5-MoE-A2.7B-Chat", "relation": "quantized", "source": "Repository name, pinned quantized config, public base model card, public base config comparison, and tensor geometry", "config_compatible": true, "notes": "The target repo name identifies Qwen1.5-MoE-A2.7B-Chat as the quantized base. Manual comparison against the public base chat config found no meaningful differences in audited decode geometry. The only checked-field differences are config-default representations: target mlp_only_layers is [] while the base omits/nulls it, and target sliding_window is null while both configs set use_sliding_window false." }, "architecture": { "canonical_architecture_id": "qwen1-5-moe-a2-7b-chat-w4a16", "max_context_tokens": 32768, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 8.310237312, "main_resident_weight_gb": 7.687907456, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.264654976, "routed_expert_weight_gb": 0.107054208, "routed_experts": 60, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_i64_bf16", "traffic_scope": "ordinary text decode through Qwen2-MoE decoder layers, model.norm, and lm_head, excluding resident-only input embedding and using expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "Qwen2MoeSparseMoeBlock computes the shared expert and shared expert gate on every sparse MoE layer. The compressed-tensors recipe leaves shared_expert_gate uncompressed but quantizes shared_expert Linear weights, and both are included in fixed_weight_gb.", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 int4 payload tensors, BF16 scales and ignored-module tensors, and tiny I64 weight_shape tensors. Routed expert tensors are byte-uniform across all 60 expert indexes, including weight_packed, weight_scale, and weight_shape side tensors." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The pinned target config records 24 layers, 16 KV heads, 128 head dimension, 32768 max position embeddings, and use_sliding_window false. Bounds Engine v1 therefore charges expanded BF16 K/V cache for all 24 layers." }, "notes": "Qwen2MoeForCausalLM chat profile using the served compressed-tensors config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, GPTQ dequantization, activation traffic, and router compute are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5761895970237549, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-gptq-w4a16-qwen2-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 int4 weights, BF16 scales, unquantized BF16 tensors, and I64 shape side tensors from safetensors headers. GPTQ dequantization, activation traffic, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 128, symmetric quantization, BF16 runtime dtype, and kv_cache_scheme null. weight_bytes_per_param records exact resident stored bytes divided by the API logical safetensors parameter total; exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "NM Testing Qwen1.5 MoE A2.7B Chat W4A16 API metadata", "url": "https://huggingface.co/api/models/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16", "source_type": "model_card", "supports": [ "repo", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 44f1710adc798a746a42b8f43165bb8efe9a71f4, the live API records a public non-gated repo with safetensors, qwen2_moe, compressed-tensors, and region:us tags. Current downloads are 143934. The target repo has no README/cardData, no pipeline_tag, and no library_name. The API safetensors block reports I64: 8976, I32: 13690208256, BF16: 732530688, and total: 14422747920 logical/storage-accounting parameters." }, { "label": "NM Testing Qwen1.5 MoE A2.7B Chat W4A16 config", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/raw/44f1710adc798a746a42b8f43165bb8efe9a71f4/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The pinned config records Qwen2MoeForCausalLM, qwen2_moe, BF16 runtime dtype, hidden size 2048, 24 layers, 16 attention heads, 16 KV heads, derived head_dim 128, 60 routed experts, 4 experts per token, shared_expert_intermediate_size 5632, decoder_sparse_step 1, empty mlp_only_layers, max_position_embeddings 32768, use_sliding_window false, tie_word_embeddings false, vocab size 151936, compressed-tensors pack-quantized INT4 weights with group size 128, ignored lm_head/router/shared_expert_gate modules, and kv_cache_scheme null." }, { "label": "Qwen1.5 MoE A2.7B Chat base model card", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B-Chat/raw/ec052fda178e241c7c443468d2fa1db6618996be/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b" ], "notes": "The base chat model card records the Tongyi Qianwen license metadata and describes Qwen1.5-MoE-A2.7B as a transformer decoder-only MoE model with 14.3B total parameters and about 2.7B activated parameters during runtime." }, { "label": "Qwen1.5 MoE A2.7B Chat base config comparison", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B-Chat/raw/ec052fda178e241c7c443468d2fa1db6618996be/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching audited architecture fields between the target config and the public base chat config: architecture, model_type, hidden size, intermediate sizes, layer count, attention/KV heads, max positions, expert count, experts per token, shared expert size, decoder_sparse_step, tied embeddings, vocab size, and use_sliding_window false. Target mlp_only_layers [] matches the base default sparse-all-layers behavior, and target sliding_window null does not affect decode because sliding-window attention is disabled." }, { "label": "NM Testing Qwen1.5 MoE A2.7B Chat W4A16 recipe", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/raw/44f1710adc798a746a42b8f43165bb8efe9a71f4/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format" ], "notes": "The recipe records a GPTQModifier W4A16 scheme targeting Linear modules while ignoring lm_head, mlp.gate, and mlp.shared_expert_gate. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "NM Testing Qwen1.5 MoE A2.7B Chat W4A16 safetensors index and shard headers", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/resolve/44f1710adc798a746a42b8f43165bb8efe9a71f4/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 8.310237312 GB across two shards. Direct range-read safetensors headers found 13635 tensors totaling exactly 8.310237312 GB, split into 6.845104128 GB I32, 1.465061376 GB BF16, and 0.000071808 GB I64. Linked shard sizes total 8.311911312 GB, leaving 0.001674000 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight contributes 0.622329856 GB resident-only. Ordinary text resident bytes therefore sum to 7.687907456 GB. Routed expert tensors sum to 6.423252480 GB and divide exactly into 60 uniform expert indexes of 0.107054208 GB. Fixed ordinary text traffic, including self-attention, routers, shared experts, shared expert gates, norms, and lm_head, sums to 1.264654976 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, pinned served compressed-tensors config, public base chat model card, public base chat config comparison, quantization recipe, safetensors index, linked-object metadata, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated ideal flat 4-bit MoE estimate, which undercounted compressed-tensors side tensors and missed the exact resident-only embedding, fixed shared expert traffic, and per-expert routed byte groups." }, { "id": "nm-testing--qwen3-coder-30b-a3b-instruct-w4a16-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq", "title": "NM Testing Qwen3-Coder 30B A3B Instruct W4A16 AWQ", "summary": "Audited memory-side text-decode bounds profile for the nm-testing compressed-tensors W4A16 AWQ package of Qwen3-Coder 30B A3B Instruct.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served compressed-tensors config, official BF16 base config comparison, recipe.yaml, and safetensors header review", "config_compatible": false, "notes": "The repo metadata records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the core ordinary-decode topology used by this profile: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, 262144 max positions, and disabled sliding-window attention. Manual comparison against the audited official BF16 base config found served-config differences in intermediate_size, max_window_layers, router_aux_loss_coef, and quantization metadata, so this profile uses the nm-testing served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-w4a16-awq-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.686185472, "main_resident_weight_gb": 16.063855616, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.115061248, "routed_expert_weight_gb": 0.116787456, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_i32_bf16_i64", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The served config records shared_expert_intermediate_size 0. Router/gate tensors are ignored by quantization and included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived bytes are used because the W4A16 package stores packed I32 int4 tensors plus BF16 side tensors and tiny I64 shape tensors. Routed expert tensors are stored as per-expert down_proj, gate_proj, and up_proj weight_packed, weight_scale, and weight_shape tensors across all 48 layers; all 128 expert indexes are byte-uniform." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, and the compressed-tensors quantization config records kv_cache_scheme null, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM W4A16 compressed-tensors profile using the served nm-testing config, recipe, and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, dequantization, activation traffic, and expert-parallel communication are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llm-compressor-compressed-tensors-w4a16-qwen3-coder-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors payload bytes: packed I32 int4 weights, BF16 scales and ignored modules, and tiny I64 shape side tensors from safetensors headers. Dequantization, activation traffic, kernel choice, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group size 128, symmetric group quantization, and kv_cache_scheme null. W4A16 and top-level torch_dtype bfloat16 imply BF16 activation/KV serving for this profile." }, "evidence": [ { "label": "NM Testing Qwen3-Coder 30B A3B W4A16 AWQ API metadata", "url": "https://huggingface.co/api/models/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "serving", "total_params_b" ], "notes": "At commit 9971cd6828ce3eefdcd9e9ca72dc4586ede07379, the live API records a public non-gated repo with safetensors, qwen3_moe, compressed-tensors, region:us, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 31877, below the older scraped qualifying catalog count of 115566 retained in the local data. The API card_data is empty and does not expose a license or pipeline tag. The API safetensors block records I32 29896998912, BF16 868694016, I64 37248, and an inconsistent total 4605856128, so exact memory figures in this profile come from direct safetensors header reads." }, { "label": "NM Testing Qwen3-Coder 30B A3B W4A16 AWQ model card", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/raw/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, says generation requires a vLLM llm-compressor change, and shows an AWQModifier recipe with duo_scaling false, W4A16, Linear targets, and ignored lm_head, mlp.gate, shared_expert_gate, and visual modules. It also reports HumanEval/HumanEval+ average score 91.93 for this package versus 91.35 for the base." }, { "label": "NM Testing Qwen3-Coder 30B A3B W4A16 AWQ recipe", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/raw/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/recipe.yaml", "source_type": "config", "supports": [ "serving", "weight_format", "fixed_weight_gb" ], "notes": "The pinned recipe uses AWQModifier targets [Linear], scheme W4A16, duo_scaling false, and the ignore list [lm_head, re:.*mlp.gate$, re:.*mlp.shared_expert_gate$, re:visual.*]. This matches the served config ignore list and supports treating lm_head and router gate tensors as unquantized fixed ordinary-decode traffic." }, { "label": "NM Testing Qwen3-Coder 30B A3B W4A16 AWQ config", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/raw/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, qkv_bias false, use_qk_norm true, and compressed-tensors pack-quantized INT4 weights with group size 128, symmetric group strategy, kv_cache_scheme null, and quantization_status compressed." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked the audited architecture fields. The nm-testing config matches the official BF16 base on the MoE routing and KV-bearing decode fields used by this profile, including layer count, heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, max_window_layers 28 vs 48, router_aux_loss_coef 0.0 vs 0.001, and quantization metadata, so config_compatible is false and fixed_weight_gb is derived from the nm-testing tensor headers." }, { "label": "NM Testing Qwen3-Coder 30B A3B W4A16 AWQ safetensors index and shard headers", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/resolve/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 16686185472 bytes across four shards. Direct range-read safetensors headers found 56115 tensors with payload bytes exactly matching the index total: 16.686185472 GB, split into 14.948499456 GB I32 packed tensors, 1.737388032 GB BF16 tensors, and 0.000297984 GB I64 weight_shape tensors. Linked shard sizes total 16.693079328 GB, leaving 0.006893856 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately and remains in fixed decode traffic. Routed expert tensors sum to 14.948794368 GB and divide exactly into 128 uniform expert groups of 0.116787456 GB. Non-expert fixed decode tensors including lm_head.weight, router/gate tensors, norms, attention, and dense side tensors sum to 1.115061248 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, pinned model card, pinned recipe, pinned served compressed-tensors config, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated 0.5-byte metadata estimate for this W4A16 repo. It uses exact stored compressed-tensors payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "nm-testing--smollm-1-7b-instruct-quantized-w4a16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16", "title": "NM Testing SmolLM 1.7B Instruct W4A16", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors W4A16 SmolLM 1.7B Instruct artifact.", "model_family": "smollm-1.7b-dense", "base_model_proof": { "base_model": "HuggingFaceTB/SmolLM-1.7B-Instruct", "relation": "quantized", "source": "Repository name, pinned quantized config, public base config comparison, and tensor geometry", "config_compatible": true, "notes": "The model card text incorrectly names SmolLM-135M in several places, but the repo id, HF API parameter count, pinned config geometry, and public base config comparison all identify this as a quantized SmolLM-1.7B-Instruct artifact. Manual comparison found no differences in the audited Llama text architecture fields between the quantized config and HuggingFaceTB/SmolLM-1.7B-Instruct." }, "architecture": { "canonical_architecture_id": "smollm-1.7b", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.83720584, "swept_params_b": 1.736542544, "auxiliary_resident_params_b": 0.100663296, "resident_weight_gb": 1.711680128, "swept_weight_gb": 1.309026944, "auxiliary_resident_weight_gb": 0.402653184, "resident_parameter_scope": "safetensors_header_logical_int4_f32_i64", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight and includes quantized layer tensors, norms, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix after token embeddings are available", "notes": "Compressed-tensors stores packed I32 int4 weights, F32 scale tensors, tiny I64 weight_shape side tensors, and unquantized F32 embed_tokens and lm_head tensors. The config records tie_word_embeddings true, but the safetensors file stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges lm_head.weight as swept output-projection traffic and treats model.embed_tokens.weight as resident-only for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 32, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 24 layers, 32 attention heads, 32 KV heads, 64 head dimension, and 2048 max position embeddings. There is no sliding-window or grouped-query reduction." }, "notes": "Dense LlamaForCausalLM profile using the served compressed-tensors config and exact stored safetensors bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.9316757495175391, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "sparseml-compressed-tensors-w4a16-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed I32 int4 weights, F32 scales, F32 ignored modules, and I64 shape side tensors from the safetensors header. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors pack-quantized INT4 weights with group size 64 and kv_cache_scheme null. The model card evaluation command uses dtype=bfloat16 and W4A16 implies 16-bit activation/KV serving, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "SmolLM 1.7B Instruct W4A16 API metadata", "url": "https://huggingface.co/api/models/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b", "revision" ], "notes": "At commit 1c917fc65b6cf47c20a2c2b4d6e2a8148fc20e9e, the API records an Apache-2.0 text-generation repo with llama, safetensors, compressed-tensors, and region:us tags. Current downloads are 1061745. The safetensors block reports logical parameters I64: 336, F32: 226592768, I32: 1610612736, and total: 1837205840." }, { "label": "SmolLM 1.7B Instruct W4A16 config", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16/raw/1c917fc65b6cf47c20a2c2b4d6e2a8148fc20e9e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, model_type llama, hidden size 2048, intermediate size 8192, 24 layers, 32 attention heads, 32 KV heads, 2048 max position embeddings, rope_theta 10000, vocab size 49152, tie_word_embeddings true, compressed-tensors pack-quantized INT4 weights with group size 64, lm_head ignored by quantization, kv_cache_scheme null, and quantization_status frozen. The top-level torch_dtype is float32, matching the F32 unquantized tensors stored in the safetensors file." }, { "label": "SmolLM 1.7B Instruct W4A16 model card", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16", "source_type": "model_card", "supports": [ "serving", "weight_format", "license" ], "notes": "The card metadata records Apache-2.0 and text-generation. The body describes INT4 weight quantization with GPTQ, group size 64, W4A16 serving, and a sparseml evaluation command using dtype=bfloat16. The body also contains stale SmolLM-135M labels, so the profile does not use those labels as architecture evidence." }, { "label": "HuggingFaceTB SmolLM 1.7B Instruct base config", "url": "https://huggingface.co/HuggingFaceTB/SmolLM-1.7B-Instruct/raw/69f49d9c36434d6a3f319dadc5bd3b812752b98b/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited architecture fields between the base config and the quantized config: LlamaForCausalLM, hidden size 2048, intermediate size 8192, 24 layers, 32 attention heads, 32 KV heads, 2048 max context, vocab size 49152, rope_theta 10000, and tied embeddings. The only audited-field dtype difference is base torch_dtype bfloat16 versus quantized torch_dtype float32." }, { "label": "SmolLM 1.7B Instruct W4A16 safetensors header", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16/resolve/1c917fc65b6cf47c20a2c2b4d6e2a8148fc20e9e/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The single safetensors file reports Content-Length 1.711742704 GB. A range-read of the 62568-byte header found 555 tensors totaling 1.711680128 GB of tensor payload: 0.805306368 GB packed I32 tensors, 0.906371072 GB F32 tensors, and 0.000002688 GB I64 tensors. Logical safetensors parameters total 1.837205840B. model.embed_tokens.weight is F32 with shape [49152, 2048] and contributes 0.100663296B parameters / 0.402653184 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Quantized layer tensors plus norms and lm_head total 1.736542544B logical parameters / 1.309026944 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, pinned quantized config, pinned model card, public base config comparison, and direct range-read safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which used an ideal flat 4-bit resident and active weight estimate and missed F32 unquantized lm_head/embed tensors plus F32 scale overhead." }, { "id": "nvidia--cosmos-reason2-2b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/Cosmos-Reason2-2B", "title": "NVIDIA Cosmos Reason2 2B BF16", "summary": "Unsupported profile stub for the gated NVIDIA Cosmos Reason2 2B repo.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-2B-Instruct", "relation": "finetune", "source": "Hugging Face API base_model tags and local scrape row; raw Cosmos config is gated in this audit environment", "config_compatible": false, "notes": "The public API records base_model Qwen/Qwen3-VL-2B-Instruct and base_model:finetune tags, but the served Cosmos config could not be downloaded with the configured osolmaz HF CLI identity. Do not assume exact base geometry, context settings, tied embedding layout, or swept traffic until the gated config and tensor header are directly accessible." }, "architecture": { "canonical_architecture_id": "cosmos-reason2-2b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.43869696, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 2438696960 BF16 safetensors parameters for this repo. Resident bytes can be estimated from public metadata, but swept decode traffic and auxiliary resident splits are not audited because the config and tensor headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is gated auto approval, and raw config access plus hf download config.json with the configured osolmaz CLI identity returned access denied in this audit environment.", "notes": "Do not infer Qwen3-VL KV heads, context length, tied embeddings, visual tower layout, or swept decode traffic from the base model tag. Replace this with an audited adapter only after direct Cosmos config and safetensors header evidence is available." }, "notes": "This profile intentionally fails closed until the gated Cosmos config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "NVIDIA Cosmos Reason2 2B API metadata", "url": "https://huggingface.co/api/models/nvidia/Cosmos-Reason2-2B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 9ce19a195e423419c349abfc86fd07178b230561, the API reports gated: auto, image-text-to-text pipeline, library cosmos, NVIDIA Open Model License, qwen3_vl and cosmos tags, base_model Qwen/Qwen3-VL-2B-Instruct, base_model:finetune tag, current downloads 672648, and safetensors BF16 count 2438696960." }, { "label": "Gated config access check", "url": "https://huggingface.co/nvidia/Cosmos-Reason2-2B/raw/9ce19a195e423419c349abfc86fd07178b230561/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config.json returned 401 Unauthorized. `hf download nvidia/Cosmos-Reason2-2B config.json --revision 9ce19a195e423419c349abfc86fd07178b230561` with the configured osolmaz CLI identity returned: Access denied. This repository requires approval." }, { "label": "Qwen3 VL 2B Instruct base profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "An audited profile exists for the public base model, but this Cosmos profile does not inherit its production bounds because the gated finetune config and tensor header were not directly verified." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context settings, tied embeddings, resident auxiliary split, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct Cosmos config and safetensors header evidence is available." }, { "id": "nvidia--cosmos-reason2-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/Cosmos-Reason2-8B", "title": "NVIDIA Cosmos Reason2 8B BF16", "summary": "Unsupported profile stub for the gated NVIDIA Cosmos Reason2 8B repo.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-8B-Instruct", "relation": "finetune", "source": "Hugging Face API base_model tags and local scrape row; raw Cosmos config is gated in this audit environment", "config_compatible": false, "notes": "The public API records base_model Qwen/Qwen3-VL-8B-Instruct and base_model:finetune tags, but the served Cosmos config could not be downloaded with the configured osolmaz HF CLI identity. Do not assume exact base geometry, context settings, tied embedding layout, or swept traffic until the gated config and tensor header are directly accessible." }, "architecture": { "canonical_architecture_id": "cosmos-reason2-8b", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 8.767123696, "parameter_scope": "hf_api_safetensors_total", "notes": "The Hugging Face API records 8767123696 BF16 safetensors parameters for this repo. Resident bytes can be estimated from public metadata, but swept decode traffic, text/vision split, and auxiliary resident components are not audited because the config and tensor headers are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is gated auto approval, and raw config/index/preprocessor access plus hf download config.json with the configured osolmaz CLI identity returned access denied in this audit environment.", "notes": "Do not infer Qwen3-VL KV heads, context length, tied embeddings, visual tower layout, or swept decode traffic from the base model tag. Replace this with an audited adapter only after direct Cosmos config and safetensors header evidence is available." }, "notes": "This profile intentionally fails closed until the gated Cosmos config and tensor headers can be audited." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because KV geometry and swept tensor traffic are unavailable.", "notes": "BF16 weight dtype comes from the Hugging Face API safetensors metadata." }, "evidence": [ { "label": "NVIDIA Cosmos Reason2 8B API metadata", "url": "https://huggingface.co/api/models/nvidia/Cosmos-Reason2-8B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit a9fae2cf89dc64db96b12860417f0eb403013bb9, the API reports gated: auto, image-text-to-text pipeline, library cosmos, NVIDIA Open Model License, qwen3_vl and cosmos tags, base_model Qwen/Qwen3-VL-8B-Instruct, base_model:finetune tag, current downloads 183046, and safetensors BF16 count 8767123696." }, { "label": "Gated config access check", "url": "https://huggingface.co/nvidia/Cosmos-Reason2-8B/raw/a9fae2cf89dc64db96b12860417f0eb403013bb9/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw config.json, model.safetensors.index.json, README.md, generation_config.json, preprocessor_config.json, and video_preprocessor_config.json returned 401 restricted-access responses. `hf download nvidia/Cosmos-Reason2-8B config.json` with the configured osolmaz CLI identity also failed." }, { "label": "Qwen3 VL 8B Instruct base profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "An audited profile exists for the public base model, but this Cosmos profile does not inherit its production bounds because the gated finetune config and tensor header were not directly verified." } ], "unsupported_reason": "Gated config and tensor bytes are not accessible in this audit environment, so KV geometry, exact context settings, tied embeddings, text/vision resident split, and swept decode traffic cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after direct Cosmos config and safetensors header evidence is available." }, { "id": "nvidia--deepseek-r1-0528-nvfp4-v2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/DeepSeek-R1-0528-NVFP4-v2", "title": "NVIDIA DeepSeek R1 0528 NVFP4 v2", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's TensorRT Model Optimizer NVFP4 DeepSeek R1 0528 v2 artifact.", "model_family": "deepseek-r1-0528-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-R1-0528", "relation": "quantized", "source": "Hugging Face card metadata, served config comparison, TensorRT Model Optimizer quantization config, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The NVIDIA config matches the DeepSeek R1 0528 base config for ordinary architecture fields. The base repo's FP8 quantization_config is absent because NVIDIA supplies ModelOpt NVFP4 details in hf_quant_config.json." }, "architecture": { "canonical_architecture_id": "deepseek-r1-0528", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 413.306995616, "main_resident_weight_gb": 384.526784416, "auxiliary_resident_weight_gb": 28.7802112, "fixed_weight_gb": 16.70274448, "routed_expert_weight_gb": 1.436812656, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model layers 0-60, model.norm.weight, and lm_head.weight, excluding full input embedding lookup and resident-only next-token prediction layer 61", "auxiliary_scope": "model.embed_tokens.weight and model.layers.61 tensors are resident in the checkpoint but are not swept as full tensors for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the artifact mixes packed U8 NVFP4 payloads, F8_E4M3 scale tensors, BF16 tensors, and small F32 scale tensors. Routed expert tensor groups are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 128, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 128, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all ordinary decoder layers." } ], "notes": "The shipped custom DeepSeek implementation computes compressed_kv internally but constructs expanded key_states and value_states before past_key_value.update. The NVIDIA hf_quant_config records FP8 KV cache storage, so Bounds Engine v1 charges expanded FP8 K/V cache streams for this serving artifact." }, "notes": "The profile models ordinary text decode for the ModelOpt v2 package. Layer 61 is the next-token-prediction auxiliary layer and remains resident-only for the ordinary generation path." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 1, "kv_store_format": "fp8_expanded_key_value_cache", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_expanded_key_value_cache", "kv_read_bytes_per_scalar": 1, "runtime_format": "tensorrt-llm-modelopt-nvfp4-fp8-kv-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed NVFP4 payloads, F8 scale tensors, BF16/F32 fixed tensors, and FP8 expanded K/V cache bytes. Dequantization, activation traffic, TensorRT-LLM scheduling, and compute overhead are outside Bounds Engine v1.", "notes": "The model card targets TensorRT-LLM on NVIDIA Blackwell. hf_quant_config records ModelOpt NVFP4 weights and FP8 KV cache; exact weight bytes are taken from safetensors headers rather than derived from weight_bytes_per_param." }, "evidence": [ { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-R1-0528-NVFP4-v2", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving" ], "notes": "At commit 25a138f28f49022958b9f2d205f9b7de0cdb6e18, the API reports a public text-generation repo with MIT license, Model Optimizer library, ModelOpt/FP4/custom_code tags, base_model deepseek-ai/DeepSeek-R1-0528, region:us, 930601 downloads, and safetensors parameters split across U8 332408029184, F8_E4M3 41551003648, BF16 19673772032, and F32 15104." }, { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 model card", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "hardware" ], "notes": "The card identifies this repo as a TensorRT Model Optimizer quantization of DeepSeek R1 0528 for TensorRT-LLM on Blackwell/B200. It states that v2 additionally quantizes the attention wo module and that transformer-block linear operators are quantized to FP4." }, { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 config", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter" ], "notes": "The config records DeepseekV3ForCausalLM, 61 hidden layers, one next-token-prediction layer, hidden size 7168, 128 attention heads, 128 KV heads, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 256 routed experts, 8 experts per token, 1 shared expert, and 163840 max position embeddings." }, { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt 0.34.1.dev3, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, lm_head exclusion, all self-attention q/kv projection exclusions, and model.layers.61* exclusion." }, { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/resolve/25a138f28f49022958b9f2d205f9b7de0cdb6e18/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 163 shards. Stored tensors sum to 413.306995616 GB: U8 332.408029184 GB, F8_E4M3 41.551003648 GB, BF16 39.347544064 GB, and F32 0.00041872 GB. Main ordinary text tensors excluding model.embed_tokens.weight and layer 61 sum to 384.526784416 GB. Resident-only model.embed_tokens.weight plus layer 61 sum to 28.7802112 GB. Main routed expert tensors sum to 367.824039936 GB, exactly 1.436812656 GB per expert index. Fixed ordinary-decode traffic excluding input embedding is 16.70274448 GB." }, { "label": "NVIDIA DeepSeek R1 0528 NVFP4 v2 custom DeepSeek runtime", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 128 heads and 192 dimensions plus value_states with 128 heads and 128 dimensions before past_key_value.update. The ordinary decoder stack uses 61 layers." }, { "label": "DeepSeek R1 0528 base config comparison", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/4236a6af538feda4548eca9ab308586007567f52/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible" ], "notes": "Manual comparison against the NVIDIA config found no ordinary architecture differences. The only config-level difference is that the base FP8 repo carries a quantization_config object while the NVIDIA repo carries ModelOpt quantization in hf_quant_config.json." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, the NVIDIA model card, served config, hf_quant_config, base config comparison, range-read safetensors shard headers, and manual review of the shipped DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It supersedes the catalog estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged a mismatched KV geometry." }, { "id": "nvidia--deepseek-v4-flash-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/DeepSeek-V4-Flash-NVFP4", "title": "NVIDIA DeepSeek V4 Flash NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's ModelOpt NVFP4 DeepSeek V4 Flash serving artifact.", "model_family": "deepseek-v4-flash-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V4-Flash", "relation": "quantized", "source": "Hugging Face card metadata, served config comparison, ModelOpt quantization config, official DeepSeek V4 Flash profile/code review, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The NVIDIA config matches the official DeepSeek V4 Flash base config for all checked ordinary architecture fields. The target adds ModelOpt mixed-precision NVFP4 expert quantization metadata and preserves the CSA/HCA text geometry." }, "architecture": { "canonical_architecture_id": "deepseek-v4-flash", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 168.266793544, "main_resident_weight_gb": 163.613944028, "auxiliary_resident_weight_gb": 4.652849516, "fixed_weight_gb": 7.786897628, "routed_expert_weight_gb": 0.6086994, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32_i8_i64", "traffic_scope": "ordinary decode through layers 0-42 plus norm.weight, head.weight, and top-level hc_head tensors, excluding resident-only embed.weight and mtp.0 tensors", "auxiliary_scope": "embed.weight and mtp.0 tensors are resident for the checkpoint but not swept for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the artifact mixes packed U8 NVFP4 expert payloads, F8_E4M3 and F8_E8M0 scale tensors, BF16 tensors, I8 tensors, I64 shape tensors, and F32 scale tensors. Routed expert tensor groups are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 43, "kv_heads": 1, "head_dim": 512, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "All main layers keep a 128-token BF16 latent KV window in the shipped DeepSeek V4 Flash inference code." }, { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00688, "notes": "Allocation coefficient for 21 ratio-4 main compressed BF16 caches, 20 ratio-128 main compressed BF16 caches, and 21 ratio-4 indexer BF16 caches." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.001504, "notes": "Read coefficient for full ratio-4 indexer scoring cache plus ratio-128 compressed cache at the default read context; the capped ratio-4 selected main-KV read is represented as a fixed read term." }, "notes": "DeepSeek V4 Flash uses Compressed Sparse Attention and Heavily Compressed Attention. Bounds Engine v1 linearizes the compressed-cache pieces from the shipped inference code." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.01220608, "read_gb_per_output_token": 0.011010048, "state_formula": "Compressor kv_state/score_state fixed buffers across 21 ratio-4 layers, 20 ratio-128 layers, and 21 ratio-4 indexer layers; fixed read term charges 512 selected compressed main-KV slots for the 21 ratio-4 indexer layers", "notes": "The allocation is true fixed compressor state. The read term is a default-workload cap approximation for the sparse top-k selected main-KV read because Bounds Engine v1 does not yet model min(index_topk, context/compression) directly." } ], "notes": "At the default 100k allocated context and 32k read context, this profile charges 0.705842176 GB allocation and 0.064774144 GB read traffic per output token for the BF16 window/compressed/indexer cache path. The ModelOpt sidecar records kv_cache_quant_algo null; the README's vLLM --kv-cache-dtype fp8 command is a runtime option, not a checkpoint-side KV scheme in the audited config." }, "notes": "The ordinary inference forward path constructs MTP modules but does not call them for ordinary text decode; MTP tensors remain resident-only in this profile." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16_window_plus_compressed_bf16_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "sparse_bf16_window_compressed_kv_plus_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "modelopt-nvfp4-deepseek-v4-csa-hca-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored NVFP4/FP8/BF16/F32/I8/I64 safetensors bytes. NVFP4 dequantization, activation quantization, sparse-attention kernel efficiency, state writes, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records ModelOpt mixed precision with NVFP4 experts, FP8 scale tensors, and kv_cache_quant_algo null. Exact weight bytes are taken from safetensors headers rather than derived from the nominal weight_bytes_per_param." }, "evidence": [ { "label": "NVIDIA DeepSeek V4 Flash NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-V4-Flash-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit e3cd60e7de98e9867116860d522499a728de1cf9, the API reports a public non-gated MIT text-generation repo with Model Optimizer library metadata, base_model deepseek-ai/DeepSeek-V4-Flash, ModelOpt/NVFP4 tags, region:us, 449002 downloads, and safetensors parameters BF16 1415259264, U8 138512695296, F8_E4M3 23337107456, I64 2327040, F8_E8M0 201694208, F32 36168018, I8 3221225472, total 166726476754." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 model card", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "total_params_b", "active_params_b", "max_context_tokens" ], "notes": "The card identifies this repo as a Model Optimizer quantization of DeepSeek-V4-Flash for SGLang and vLLM on NVIDIA Blackwell. It states 284B total parameters, 13B activated parameters, maximum context length of 1 million tokens, and that only the weights and activations of linear operators within transformer-block MoE are quantized." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention", "serving" ], "notes": "The config records DeepseekV4ForCausalLM, 43 hidden layers, one MTP layer, 256 routed experts, 6 experts per token, 1 shared expert, hidden size 4096, 64 attention heads, 1 KV head, head_dim 512, 1048576 max position embeddings, sliding_window 128, index_head_dim 128, index_topk 512, expert_dtype fp4, ModelOpt mixed precision, moe_quant_algo NVFP4, kv_cache_quant_algo null, and compress_ratios with two uncompressed layers, 21 ratio-4 layers, 20 ratio-128 layers, and an uncalled MTP ratio entry." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt dsv4-nvfp4-experts, quant_algo MIXED_PRECISION, kv_cache_quant_algo null, group_size 16, every layers.*.ffn.experts module quantized with NVFP4, and exclusions for attention modules, shared experts, head, and mtp." }, { "label": "DeepSeek V4 Flash base config comparison", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "kv_adapter" ], "notes": "Manual comparison found no differences across 24 checked profile-relevant fields between the NVIDIA NVFP4 config and the official DeepSeek V4 Flash config: architecture, model type, hidden layer counts, MTP layer count, hidden size, MoE geometry, attention geometry, CSA/HCA index fields, sliding window, max context, tied embeddings, vocabulary, expert dtype, and compress_ratios. The only config-level difference is quantization_config." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/resolve/e3cd60e7de98e9867116860d522499a728de1cf9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "Safetensors headers were range-read across all 46 shards. Stored tensors sum to the index metadata.total_size, 168.266793544 GB: U8 138.512695296 GB, F8_E4M3 23.337107456 GB, BF16 2.830518528 GB, I8 3.221225472 GB, F8_E8M0 0.201694208 GB, F32 0.144936264 GB, and I64 0.018616320 GB. Linked-object sizes total 168.281985176 GB, leaving 0.015191632 GB of safetensors header/container overhead outside tensor payloads. Ordinary swept tensors under layers plus norm, head, and top-level hc_head tensors sum to 163.613944028 GB. Resident-only embed.weight plus mtp.0 tensors sum to 4.652849516 GB. Routed expert tensors sum to 155.827046400 GB, exactly 0.608699400 GB per expert index. Fixed ordinary-decode traffic including shared experts sums to 7.786897628 GB." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 inference code", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/inference/model.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "compressed_attention", "kv_adapter", "mtp_resident_only" ], "notes": "Manual review found Transformer.forward loops over the 43 main layers, applies norm and head, and does not call self.mtp in ordinary generation. Attention keeps a 128-token window, optional compressed cache slots according to compress_ratio, and an Indexer cache for ratio-4 layers. The code comments that indexer compressed KV could use FP8 but the current implementation uses BF16." }, { "label": "NVIDIA DeepSeek V4 Flash NVFP4 inference config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/inference/config.json", "source_type": "config", "supports": [ "serving", "kv_adapter", "weight_format" ], "notes": "The inference config records dtype fp8, expert_dtype fp4, scale_fmt ue8m0, 43 layers, 6 activated experts, 128-token window, index_topk 512, index_head_dim 128, and the same compression-ratio schedule used for the KV adapter arithmetic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, NVIDIA model card, served config, hf_quant_config, base config comparison, NVIDIA inference config/code, safetensors index, linked-object metadata, and direct range-read safetensors header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It supersedes the scraped metadata estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged a generic full-context KV cache." }, { "id": "nvidia--deepseek-v4-pro-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/DeepSeek-V4-Pro-NVFP4", "title": "NVIDIA DeepSeek V4 Pro NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's ModelOpt NVFP4 DeepSeek V4 Pro serving artifact.", "model_family": "deepseek-v4-pro-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V4-Pro", "relation": "quantized", "source": "Hugging Face card metadata, served config comparison, ModelOpt quantization config, audited official DeepSeek V4 Pro profile, NVIDIA inference config/code, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The NVIDIA config preserves the official DeepSeek V4 Pro CSA/HCA text geometry: 61 main layers, one MTP layer, 384 routed experts, 6 experts per token, 1 shared expert, 128-token window, 1024 index top-k, and the same compression-ratio schedule. The target adds ModelOpt mixed-precision NVFP4 expert quantization metadata and different stored tensor encodings." }, "architecture": { "canonical_architecture_id": "deepseek-v4-pro", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 913.061485688, "main_resident_weight_gb": 897.251539916, "auxiliary_resident_weight_gb": 15.809945772, "fixed_weight_gb": 26.84016428, "routed_expert_weight_gb": 2.266695096, "routed_experts": 384, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32_i8_i64", "traffic_scope": "ordinary decode through layers 0-60 plus norm.weight, head.weight, and top-level hc_head tensors, excluding resident-only embed.weight and mtp.0 tensors", "auxiliary_scope": "embed.weight and mtp.0 tensors are resident for the checkpoint but not swept for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because this ModelOpt package mixes packed U8 NVFP4 expert payloads, FP8 tensors/scales, BF16 tensors, I8 tensors, I64 shape tensors, and F32 scale tensors. Routed expert tensor groups are byte-uniform across all 384 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 61, "kv_heads": 1, "head_dim": 512, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "All main layers keep a 128-token BF16 latent KV window in the shipped DeepSeek V4 Pro inference code." }, { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.009848, "notes": "Allocation coefficient for 30 ratio-4 main compressed BF16 caches, 31 ratio-128 main compressed BF16 caches, and 30 ratio-4 indexer BF16 caches." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.002168, "notes": "Read coefficient for full ratio-4 indexer scoring cache plus ratio-128 compressed cache at the default read context; the capped ratio-4 selected main-KV read is represented as a fixed read term." }, "notes": "DeepSeek V4 Pro uses Compressed Sparse Attention and Heavily Compressed Attention. Bounds Engine v1 linearizes the compressed-cache pieces from the official/NVIDIA inference code." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.018710528, "read_gb_per_output_token": 0.03145728, "state_formula": "Compressor kv_state/score_state fixed buffers across 30 ratio-4 layers, 31 ratio-128 layers, and 30 ratio-4 indexer layers; fixed read term charges 1024 selected compressed main-KV slots for the 30 ratio-4 indexer layers", "notes": "The allocation is true fixed compressor state. The read term is a default-workload cap approximation for the sparse top-k selected main-KV read because Bounds Engine v1 does not yet model min(index_topk, context/compression) directly." } ], "notes": "At the default 100k allocated context and 32k read context, this profile charges 1.01150592 GB allocation and 0.108828672 GB read traffic per output token for the BF16 window/compressed/indexer cache path. The ModelOpt sidecar records kv_cache_quant_algo null." }, "notes": "The ordinary inference forward path constructs MTP modules but does not call them for ordinary text decode; MTP tensors remain resident-only in this profile." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16_window_plus_compressed_bf16_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "sparse_bf16_window_compressed_kv_plus_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "modelopt-nvfp4-deepseek-v4-pro-csa-hca-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored NVFP4/FP8/BF16/F32/I8/I64 safetensors bytes. NVFP4 dequantization, activation quantization, sparse-attention kernel efficiency, state writes, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records ModelOpt mixed precision with NVFP4 experts, FP8 scale tensors, and kv_cache_quant_algo null. Exact weight bytes are taken from safetensors headers rather than derived from the nominal weight_bytes_per_param." }, "evidence": [ { "label": "NVIDIA DeepSeek V4 Pro NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-V4-Pro-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 1449d1e641023406daf6b432361486c768aad740, the API reports a public non-gated MIT text-generation repo with Model Optimizer library metadata, base_model deepseek-ai/DeepSeek-V4-Pro, ModelOpt/NVFP4 tags, region:us, 152280 downloads, and safetensors parameters BF16 2816899328, I64 2327040, F32 87776414, F8_E8M0 794137600, F8_E4M3 119881596928, U8 773698093056, I8 12683575296, total 909964405662." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 model card", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "total_params_b", "active_params_b", "max_context_tokens" ], "notes": "The card identifies this repo as a Model Optimizer quantization of DeepSeek-V4-Pro for SGLang and vLLM on NVIDIA Blackwell/GB300. It states 1.6T total parameters, 49B activated parameters, maximum context length of 1 million tokens, and that only the weights and activations of linear operators within transformer-block MoE are quantized." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention", "serving" ], "notes": "The config records DeepseekV4ForCausalLM, 61 hidden layers, one MTP layer, 384 routed experts, 6 experts per token, 1 shared expert, hidden size 7168, 128 attention heads, 1 KV head, head_dim 512, 1048576 max position embeddings, sliding_window 128, index_head_dim 128, index_topk 1024, expert_dtype fp4, ModelOpt mixed precision, moe_quant_algo NVFP4, kv_cache_quant_algo null, and the same Pro compression-ratio schedule as the audited base." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt dsv4-nvfp4-experts, quant_algo MIXED_PRECISION, kv_cache_quant_algo null, group_size 16, MoE expert modules quantized with NVFP4, and exclusions for attention modules, shared experts, head, and mtp." }, { "label": "DeepSeek V4 Pro audited base profile", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "config_compatible", "kv_adapter" ], "notes": "The existing audited official Pro profile and config establish the same 61-layer CSA/HCA geometry, ordinary decode scope, MTP resident-only treatment, and BF16 compressed/indexer KV state formulas used here. The NVIDIA profile changes stored tensor encodings and per-expert bytes, not the text decode topology." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/resolve/1449d1e641023406daf6b432361486c768aad740/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "Safetensors headers were range-read across all 64 shards. Stored tensors sum to the index metadata.total_size, 913.061485688 GB: U8 773.698093056 GB, F8_E4M3 119.881596928 GB, I8 12.683575296 GB, BF16 5.633798656 GB, F8_E8M0 0.794137600 GB, F32 0.351667832 GB, and I64 0.018616320 GB. Linked-object sizes total 913.092561960 GB, leaving 0.031076272 GB of safetensors header/container overhead outside tensor payloads. Ordinary swept tensors under layers plus norm, head, and top-level hc_head tensors sum to 897.251539916 GB. Resident-only embed.weight plus mtp.0 tensors sum to 15.809945772 GB. Routed expert tensors sum to 870.410916864 GB, exactly 2.266695096 GB per expert index. Fixed ordinary-decode traffic including shared experts sums to 26.840164280 GB." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 inference code", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/inference/model.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "compressed_attention", "kv_adapter", "mtp_resident_only" ], "notes": "Manual review matched the official Pro ordinary path: Transformer.forward loops over the 61 main layers, applies norm and head, and does not call self.mtp in ordinary generation. Attention keeps a 128-token window, optional compressed cache slots according to compress_ratio, and an Indexer cache for ratio-4 layers." }, { "label": "NVIDIA DeepSeek V4 Pro NVFP4 inference config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/inference/config.json", "source_type": "config", "supports": [ "serving", "kv_adapter", "weight_format" ], "notes": "The NVIDIA inference config records dtype fp8, expert_dtype fp4, scale_fmt ue8m0, 61 layers, 6 activated experts, 128-token window, index_topk 1024, index_head_dim 128, and the same compression-ratio schedule used for the KV adapter arithmetic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, NVIDIA model card, served config, hf_quant_config, official DeepSeek V4 Pro profile/config comparison, NVIDIA inference config/code, safetensors index, linked-object metadata, and direct range-read safetensors header byte grouping." }, "notes": "This profile is expected to be resident_not_fit on current local single-machine hardware rows, including 128GB local systems and the 748GB GB300-class row. It supersedes the scraped metadata estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged generic full-context KV cache." }, { "id": "nvidia--diffusiongemma-26b-a4b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/diffusiongemma-26B-A4B-it-NVFP4", "title": "NVIDIA DiffusionGemma 26B A4B IT NVFP4", "summary": "Unsupported profile stub with exact resident tensor evidence for the NVIDIA ModelOpt NVFP4 DiffusionGemma serving artifact.", "model_family": "diffusion-gemma-block-diffusion-moe", "base_model_proof": { "base_model": "google/diffusiongemma-26B-A4B-it", "relation": "quantized", "source": "Hugging Face cardData base_model metadata, base_model tags, NVIDIA model card, served config, and ModelOpt quantization config", "config_compatible": true, "notes": "The served config preserves the DiffusionGemmaForBlockDiffusion architecture, 256-token canvas, 30 decoder layers, hybrid local/global attention, 128 experts, top_k_experts 8, and 262144 max position embeddings while adding ModelOpt NVFP4 weight and FP8 KV-cache quantization metadata." }, "architecture": { "canonical_architecture_id": "diffusion-gemma-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.818050776, "main_resident_weight_gb": 17.672461884, "auxiliary_resident_weight_gb": 1.145588892, "fixed_weight_gb": 4.826003004, "routed_expert_weight_gb": 0.10036296, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f8_e4m3_u8_f32", "traffic_scope": "Exact decoder tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because DiffusionGemma block diffusion is not ordinary one-output-token autoregressive decode.", "auxiliary_scope": "model.encoder tensors are resident for the multimodal encoder/cache path but are not enough to define ordinary token-by-token swept traffic.", "shared_expert_notes": "The config records top_k_experts 8 and 128 routed experts. The checkpoint also stores dense model.decoder.layers.*.mlp.* tensors outside model.decoder.layers.*.experts.*, so those always-on/shared tensors are included in fixed_weight_gb.", "notes": "Header-derived bytes are used. model.decoder tensors total 17.672461884 GB, model.encoder tensors total 1.145588892 GB, and no lm_head tensor is stored separately. Routed expert tensors total 12.84645888 GB and divide exactly into 128 uniform expert groups of 0.10036296 GB." }, "kv_adapter": { "kind": "unknown", "reason": "DiffusionGemma uses block diffusion over a 256-token canvas with a decoder that applies bidirectional attention over the generation canvas and then appends fully denoised canvases to cache. Bounds Engine v1 only models ordinary autoregressive per-output-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "The ModelOpt files record FP8 KV-cache quantization, but that only identifies scalar storage. A production profile still needs a dedicated block-diffusion adapter with canvas length, denoising iteration count, sampler behavior, canvas self-attention traffic, cross-attention/context-cache traffic, and block append policy. Do not reuse the Gemma 4 autoregressive KV adapter for this repo." }, "notes": "This profile intentionally fails closed even though config, quantization metadata, and tensor headers are accessible, because the supported comparison math does not model DiffusionGemma's block-diffusion generation algorithm." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-block-diffusion-vllm-modelopt-nvfp4-fp8-kv", "dequantization_notes": "The ModelOpt quantization config records NVFP4 weights and FP8 KV-cache quantization. Bounds Engine v1 does not turn those bytes into production tok/s for this repo because the generation algorithm is block diffusion rather than ordinary autoregressive decode.", "notes": "The model card documents vLLM serving, and the served config/hf_quant_config record NVFP4 weights plus an 8-bit float KV cache scheme." }, "evidence": [ { "label": "NVIDIA DiffusionGemma NVFP4 Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/diffusiongemma-26B-A4B-it-NVFP4", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 2ea837236295d617ac27f8c17d61228081932c40, the API reports a public Apache-2.0 text-generation repo with base_model google/diffusiongemma-26B-A4B-it, ModelOpt/NVFP4 tags, 1,307,197 downloads, region:us, and safetensors parameters BF16 2,985,795,948, F8_E4M3 1,427,374,080, U8 11,418,992,640, total 14,404,786,224." }, { "label": "NVIDIA DiffusionGemma NVFP4 served config", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "weight_format", "kv_store_format", "unsupported_reason" ], "notes": "The config records DiffusionGemmaForBlockDiffusion, model_type diffusion_gemma, canvas_length 256, 30 decoder layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 16 attention heads, 8 local KV heads, 2 global KV heads, 128 experts, top_k_experts 8, 262144 max position embeddings, and a static 8-bit float KV cache scheme." }, { "label": "NVIDIA DiffusionGemma NVFP4 quantization config", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "fixed_weight_gb" ], "notes": "The ModelOpt quantization config records quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes lm_head, embed_vision, every mlp/router/self_attn/self_conditioning module, and the vision tower from quantization. Direct headers confirm those excluded decoder tensors remain BF16 fixed traffic while routed expert payloads and scales are quantized." }, { "label": "NVIDIA DiffusionGemma NVFP4 model card", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm", "runtime_format" ], "notes": "The card says DiffusionGemma produces text through discrete diffusion, generates tokens in parallel 256-token blocks, uses an encoder-decoder design with bidirectional attention, supports 256K context, and is served through vLLM. It also states the artifact is an NVFP4 quantization of Gemma-26B-A4B-IT." }, { "label": "NVIDIA DiffusionGemma NVFP4 generation config", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config records max_denoising_steps 48, max_new_tokens 256, EntropyBoundSamplerConfig, confidence_threshold 0.005, stability_threshold 1, t_min 0.4, and t_max 0.8, confirming a diffusion sampling policy rather than a normal one-token decode loop." }, { "label": "NVIDIA DiffusionGemma NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 18.818050776 GB across two shards. Range-read safetensors headers found 47,067 tensors: 5.971591896 GB BF16, 1.42737408 GB F8_E4M3, 11.41899264 GB U8, and 0.00009216 GB F32. model.decoder tensors total 17.672461884 GB; model.encoder tensors total 1.145588892 GB. model.decoder.embed_tokens.weight is 1.476395008 GB, and there is no separate lm_head tensor. Non-expert decoder tensors total 4.826003004 GB. Routed expert tensors under model.decoder.layers.*.experts.* total 12.84645888 GB and divide exactly into 128 uniform expert groups of 0.10036296 GB." }, { "label": "Google DiffusionGemma 26B A4B IT base profile", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it", "source_type": "manual_review", "supports": [ "unsupported_reason", "base_model_proof" ], "notes": "The base BF16 repo is already fail-closed in Local Frontier for the same architectural reason: Bounds Engine v1 lacks a block-diffusion throughput adapter." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Reviewed from current HF API metadata, the pinned served config, hf_quant_config, generation config, model card, safetensors index, direct shard header byte grouping, and the existing base DiffusionGemma profile. Marked unsupported because Bounds Engine v1 lacks a block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model block-diffusion generation over a denoised canvas, so ordinary autoregressive throughput would be misleading even though the NVIDIA ModelOpt resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated DiffusionGemma block-diffusion adapter exists." }, { "id": "nvidia--gemma-4-26b-a4b-nvfp4", "version": "1.0.2", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Gemma-4-26B-A4B-NVFP4", "title": "NVIDIA Gemma 4 26B A4B NVFP4", "summary": "Audited memory-side bounds profile for the NVFP4 Gemma 4 26B A4B serving artifact.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and NVIDIA model card", "config_compatible": true, "notes": "The profile embeds the Gemma 4 26B A4B architecture and the NVIDIA NVFP4 serving representation." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.782360732, "main_resident_weight_gb": 17.6367719, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 4.79031302, "routed_expert_weight_gb": 0.10036296, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f8_e4m3_u8_f32", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges BF16 fixed language tensors plus expected distinct quantized routed experts", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The config records top_k_experts 8 and the public card describes 1 shared expert. ModelOpt leaves model.language_model.layers.*.mlp.*, router, attention, and embedding tensors unquantized; this profile charges those shared/always-on tensors in fixed_weight_gb.", "notes": "The original bounds note used idealized rounded 26B/4B NVFP4 traffic. This production profile uses exact stored bytes for the NVIDIA ModelOpt artifact: BF16 fixed text tensors plus quantized routed expert weights and scales." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile models the actual NVIDIA ModelOpt artifact: quantized routed experts, BF16 fixed language tensors, resident vision tensors, and FP8 KV cache with unified full-attention K/V and separate sliding K/V." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv", "dequantization_notes": "The memory-side bound charges stored tensor bytes, including quantization weights/scales, plus FP8 KV bytes; compute and dequantization overheads are outside Bounds Engine v1.", "notes": "The ModelOpt quantization config records NVFP4 weight quantization and FP8 KV-cache quantization. The served config records bfloat16 text dtype plus an 8-bit float KV cache scheme, so Bounds Engine v1 charges FP8 KV for this vLLM ModelOpt artifact." }, "evidence": [ { "label": "Google Gemma 4 26B A4B model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The config records 30 language layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, and 262144 max position embeddings." }, { "label": "NVIDIA Gemma 4 NVFP4 model card", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "weight_format", "license" ], "notes": "The card and API metadata identify the artifact as an Apache-2.0 Model Optimizer NVFP4 quantization of google/gemma-4-26B-A4B-it." }, { "label": "NVIDIA Gemma 4 NVFP4 Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Gemma-4-26B-A4B-NVFP4", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "weight_format", "total_params_b" ], "notes": "The API response at commit a19cfe00be84568a6867111c9a68c9c44fdcffe6 records safetensors parameters BF16: 2967950926, F8_E4M3: 1427374080, U8: 11418992640, and total: 14386941232." }, { "label": "NVIDIA Gemma 4 NVFP4 served config", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4/raw/main/config.json", "source_type": "config", "supports": [ "text_dtype", "kv_store_format", "kv_read_format" ], "notes": "The served config records text_config.dtype bfloat16 and quantization_config.kv_cache_scheme as static 8-bit float cache storage." }, { "label": "NVIDIA Gemma 4 NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4/raw/main/hf_quant_config.json", "source_type": "derived_calculation", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "fixed_weight_gb" ], "notes": "The ModelOpt quantization config records quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes lm_head, model.embed_vision, model.vision_tower, and every language-layer mlp/router/self_attn module from quantization. As a result, fixed ordinary text traffic remains BF16 while routed expert tensors carry the quantized payload and scales." }, { "label": "NVIDIA Gemma 4 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "tie_word_embeddings" ], "notes": "The index records total_size 18782360732 bytes across two shards. Range-read safetensors headers found 47033 tensors, no separate lm_head.weight, and stored tensors totaling 18.782360732 GB: 5.935901852 GB BF16, 1.42737408 GB F8_E4M3, 11.41899264 GB U8, and 0.00009216 GB F32. Resident text tensors under model.language_model total 17.6367719 GB; resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic, defined as all model.language_model tensors except model.language_model.layers.*.experts.* quantized payload/scale tensors, totals 4.79031302 GB. Routed expert tensors total 12.84645888 GB and divide exactly into 128 uniform expert groups of 0.10036296 GB." }, { "label": "Original local frontier bounds note", "url": "https://github.com/osolmaz/onurclaw/blob/main/docs/2026-06-30-local-frontier-model-bounds.md", "source_type": "manual_review", "supports": [ "worked_example_parameters", "bounds_regression_target" ], "notes": "The original bounds note used an idealized rounded 26B resident / 4B active NVFP4 approximation. This profile intentionally diverges from that approximation where the actual NVIDIA ModelOpt artifact stores fixed text tensors in BF16." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from Google and NVIDIA model cards, the served config, HF API metadata, ModelOpt quantization config, safetensors index, direct shard header byte grouping, and the original bounds note. FP8 KV charging is backed by hf_quant_config.json kv_cache_quant_algo FP8 and config.json quantization_config.kv_cache_scheme." }, "notes": "Production code must use this self-contained exact artifact profile rather than inferring NVFP4 serving assumptions from the base Gemma repo name or from the idealized worked example." }, { "id": "nvidia--gemma-4-31b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Gemma-4-31B-IT-NVFP4", "title": "NVIDIA Gemma 4 31B IT NVFP4", "summary": "Audited memory-side text-decode bounds profile for the NVIDIA ModelOpt NVFP4 Gemma 4 31B IT serving artifact.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, and vision geometry between the NVIDIA ModelOpt artifact and the Google BF16 base repo." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 22.169155796, "swept_params_b": 21.59341226, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 32.633255032, "swept_weight_gb": 31.48176796, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "safetensors_header_stored_bf16_f8_e4m3_u8_f32", "swept_parameter_scope": "model.language_model safetensors headers, including ModelOpt NVFP4 payload tensors, FP8 scale tensors, F32 input scales, and unquantized BF16 attention, norms, and tied embedding/output projection", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary generated text tokens", "notes": "The served config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. Header-derived bytes are used because this ModelOpt artifact stores a mixed BF16, U8, F8_E4M3, and F32 representation. Hugging Face metadata total_parameters excludes some stored side tensors, so the profile uses range-read safetensors header totals for resident and swept memory." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored tensor bytes, including ModelOpt payload and scale tensors, plus FP8 KV bytes; compute and dequantization overheads are outside Bounds Engine v1.", "notes": "The served config records quant_method modelopt, quant_algo NVFP4, text_config.dtype bfloat16, and a static 8-bit float KV cache scheme. hf_quant_config.json independently records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Gemma 4 31B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/Gemma-4-31B-IT-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit e5ef03afa233c35cb000323ff098d4291e1dd07c, the API reports a Model Optimizer text-generation repo with base_model google/gemma-4-31B-it, NVFP4/ModelOpt tags, region:us, safetensors parameters BF16: 10464098156, F8_E4M3: 1300561920, U8: 10404495360, and index total_size 32633255032 bytes." }, { "label": "NVIDIA Gemma 4 31B NVFP4 served config", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, ModelOpt NVFP4 quantization, static 8-bit float KV cache storage, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, and 262144 max position embeddings." }, { "label": "NVIDIA Gemma 4 31B NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/hf_quant_config.json", "source_type": "derived_calculation", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "swept_weight_gb" ], "notes": "The ModelOpt quantization config records producer modelopt 0.37.0, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and exclude_modules for lm_head, model.embed_vision, model.vision_tower, and every language-layer self_attn module." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this NVIDIA ModelOpt artifact; the NVIDIA artifact adds quantization_config while preserving the base architecture." }, { "label": "NVIDIA Gemma 4 31B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records total_size 32633255032 bytes across four shards. Range-read safetensors headers found 1728 tensors totaling 32.633255032 GB: 20.928196312 GB BF16, 10.40449536 GB U8, 1.30056192 GB F8_E4M3, and 0.00000144 GB F32. Language tensors under model.language_model total 21.59341226 stored tensor params / 31.48176796 GB and are swept for ordinary text decode, including ModelOpt side tensors and the tied embedding/output projection. Resident-only vision tensors under model.vision_tower plus model.embed_vision total 0.575743536 BF16 params / 1.151487072 GB. The index has no separate lm_head.weight." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the NVIDIA model card, served config, ModelOpt quantization config, base config comparison, HF API metadata, safetensors index, and direct shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal weights from per-token swept language weights and charges FP8 KV because the NVIDIA ModelOpt artifact declares FP8 KV-cache quantization." }, { "id": "nvidia--glm-5-2-nvfp4", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/GLM-5.2-NVFP4", "title": "NVIDIA GLM-5.2 NVFP4", "summary": "Audited memory-side text-decode bounds profile for the ModelOpt NVFP4 GLM-5.2 checkpoint with FP8 KV cache and IndexShare DSA.", "model_family": "glm-moe-dsa", "base_model_proof": { "base_model": "zai-org/GLM-5.2", "relation": "quantized", "source": "Hugging Face API/model-card metadata, target config, hf_quant_config, base config comparison, safetensors index/header audit, and existing GLM-5.2 profiles", "config_compatible": true, "notes": "The NVIDIA repo card and API metadata identify this package as a ModelOpt NVFP4 quantization of zai-org/GLM-5.2. The target config matches the base GLM-5.2 memory-relevant text geometry and records a static 8-bit float KV cache scheme." }, "architecture": { "canonical_architecture_id": "glm-5-2", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 464.795267072, "main_resident_weight_gb": 442.986259968, "auxiliary_resident_weight_gb": 21.809007104, "fixed_weight_gb": 35.299450368, "routed_expert_weight_gb": 1.5925266, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_modelopt_nvfp4_bf16_fp8_f32", "traffic_scope": "ordinary vLLM/SGLang text decode through layers 0-77, excluding input embedding lookup and auxiliary layer 78 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.78 tensors are resident in the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are excluded from ModelOpt quantization and included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the repo stores routed expert matrices as packed U8 plus FP8/BF16/F32 scale tensors, while attention, shared experts, dense layers, embeddings, and lm_head remain BF16/F32. Expected-distinct routing is applied to the 256 uniform routed expert indexes across sparse layers 3-77." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 192 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 256 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 21, "kv_heads": 1, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "DSA IndexShare indexer key cache component. The served config has 21 full indexer layers and 57 shared layers, so shared layers reuse top-k indices and do not run update_indexer." } ], "notes": "The GLM-5.2 DSA cache structure mirrors the audited zai-org/GLM-5.2-FP8 profile, but this NVIDIA package records kv_cache_scheme FP8 and its vLLM command passes --kv-cache-dtype fp8_e4m3, so cache scalars are charged at one byte." }, "notes": "The served config records 78 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 78 tensors are kept resident but excluded from ordinary causal decode traffic. GLM-5.2 uses IndexShare scheduling with 21 full and 57 shared ordinary indexer layers." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.6169876465760088, "kv_store_format": "fp8_e4m3_expanded_key_value_cache_plus_indexshare_dsa_indexer", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_e4m3_expanded_key_value_cache_plus_indexshare_dsa_indexer", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-sglang-modelopt-nvfp4-fp8-kv-indexshare-dsa-memory-bound", "dequantization_notes": "The memory-side bound charges stored U8 packed expert matrices, F8_E4M3 scale tensors, BF16 embeddings/attention/shared-expert tensors, and F32 scale/correction tensors from safetensors headers. Dequantization, activation traffic, sparse attention compute savings, expert-parallel placement, and speculative decoding are outside Bounds Engine v1.", "notes": "The model card's vLLM command explicitly uses --kv-cache-dtype fp8_e4m3, and config.quantization_config.kv_cache_scheme records static 8-bit float KV. The profile therefore charges FP8 cache scalars, unlike the zai-org BF16/FP8 source profiles that use BF16 KV." }, "evidence": [ { "label": "NVIDIA GLM-5.2 NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/GLM-5.2-NVFP4", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "license", "pipeline", "serving" ], "notes": "At commit aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa, the API reports a public non-gated MIT text-generation repo with safetensors, glm_moe_dsa, ModelOpt, GLM-5, fp4, modelopt, base_model zai-org/GLM-5.2, region:us, and 449881 downloads." }, { "label": "NVIDIA GLM-5.2 NVFP4 model card", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4/raw/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "runtime_format", "weight_format", "kv_adapter" ], "notes": "The card describes GLM-5.2-NVFP4 as a ModelOpt quantization of ZAI GLM-5.2, records 753B total and 40B activated parameters, says only transformer-block MoE expert Linear operators are quantized while shared experts are not, and provides SGLang/vLLM serving commands. The vLLM command includes --kv-cache-dtype fp8_e4m3." }, { "label": "NVIDIA GLM-5.2 NVFP4 config and quant config", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4/raw/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records GlmMoeDsaForCausalLM, glm_moe_dsa, dtype bfloat16, 78 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 6144, intermediate_size 12288, moe_intermediate_size 2048, 64 attention heads, 64 key/value heads, q_lora_rank 2048, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, qk_head_dim 256, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab_size 154880, max_position_embeddings 1048576, NVFP4 ModelOpt Linear quantization, and kv_cache_scheme as static 8-bit float. hf_quant_config.json confirms NVFP4, FP8 KV cache, group_size 16, and excludes lm_head, embeddings, dense layers 0-2, all self_attn modules, all shared experts, and model.layers.78." }, { "label": "zai-org GLM-5.2 base config", "url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "kv_adapter" ], "notes": "Manual comparison found the target NVFP4 config matches the audited base GLM-5.2 memory-relevant text geometry. The quantized config differs by ModelOpt quantization_config and static FP8 KV cache scheme." }, { "label": "NVIDIA GLM-5.2 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4/resolve/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 47 shards. Stored tensors sum exactly to the index total_size of 464.795267072 GB: U8 362.387865600 GB, BF16 57.108379648 GB, F8_E4M3 45.298483200 GB, and F32 0.000538624 GB across 232385 tensors. Linked shard bytes total 464.823042096 GB, leaving 0.027775024 GB of safetensors header/container overhead outside tensor payloads. The input embedding contributes 1.903165440 GB resident-only, and auxiliary layer 78 contributes 19.905841664 GB resident-only. Ordinary decode main resident tensors therefore sum to 442.986259968 GB. Routed expert tensors under model.layers.3-77 .mlp.experts.N. sum to 407.686809600 GB and divide exactly into 256 uniform expert groups of 1.592526600 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, gates, norms, routers, shared experts, and the GLM-5.2 full-indexer tensors, sums to 35.299450368 GB, so fixed plus all 256 routed expert groups reproduces the 442.986259968 GB main resident total exactly. An earlier revision of this profile recorded 1.668024072 GB per routed expert from a tensor grouping that double-counted the always-on shared-expert tensors inside the routed sum; the corrected split was re-derived from the same shard headers at the same commit." }, { "label": "Audited GLM-5.2 FP8 profile and Transformers GLM MoE DSA implementation", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "This profile reuses the audited GLM-5.2 IndexShare DSA cache structure: expanded key cache and expanded value cache for all 78 ordinary decoder layers plus DSA indexer key state for 21 full indexer layers. The difference is serving precision: this NVIDIA ModelOpt package records and documents FP8 KV cache, so each scalar is charged at one byte." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-09", "notes": "Audited from live HF API metadata, pinned model card, target config, hf_quant_config, base GLM-5.2 config comparison, direct safetensors header range reads across all 47 shards, and the existing audited GLM-5.2 DSA/IndexShare profiles. Re-reviewed 2026-07-09: the routed expert byte split was re-derived from the same pinned shard headers after a saturation consistency check flagged that fixed plus 256 routed experts exceeded the main resident total." }, "notes": "This profile models ordinary text decode with documented FP8 KV cache. It intentionally does not assume expert-parallel placement details, sparse attention compute savings, runtime-specific MLA compression, or speculative decoding speedups without direct implementation evidence." }, { "id": "nvidia--kimi-k2-5-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Kimi-K2.5-NVFP4", "title": "NVIDIA Kimi K2.5 NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 Kimi K2.5 serving artifact.", "model_family": "kimi-k2-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-K2.5", "relation": "quantized", "source": "Hugging Face model card base_model metadata, NVIDIA model card, served config, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The profile embeds the Kimi K2.5 architecture and the NVIDIA ModelOpt NVFP4/FP8-KV serving representation." }, "architecture": { "canonical_architecture_id": "kimi-k2-5", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 590.779912064, "main_resident_weight_gb": 587.488813984, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 16.727372704, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60 and lm_head, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. ModelOpt quantizes MoE linear operators but leaves self-attention, lm_head, vision_tower, and mm_projector excluded; shared expert tensors are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the artifact mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and small F32 scales. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but constructs expanded key_states and value_states before past_key_values.update. The NVIDIA config and hf_quant_config record FP8 KV cache storage, so Bounds Engine v1 charges expanded FP8 K/V cache streams for this serving artifact." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image/video encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8_expanded_key_value_cache", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_expanded_key_value_cache", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed NVFP4 payloads, F8 scale tensors, BF16/F32 fixed tensors, and FP8 expanded K/V cache bytes. Dequantization, activation traffic, vision/video prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records ModelOpt NVFP4 weights/activations with group_size 16 and a static 8-bit float KV cache scheme; hf_quant_config records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Kimi K2.5 NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Kimi-K2.5-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving" ], "notes": "At commit 0fd0a5e6879298d3476e3b61852a79792a35ae3d, the API reports a public text-generation repo with license other, Model Optimizer library, ModelOpt/NVFP4 tags, custom_code, base_model moonshotai/Kimi-K2.5, downloads 1,206,834, and usedStorage 590850735131 bytes." }, { "label": "NVIDIA Kimi K2.5 NVFP4 model card", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "model_family", "max_context_tokens" ], "notes": "The card identifies the repo as a Model Optimizer quantization of Moonshot AI Kimi K2.5, with DeepSeek V3 network architecture, 1T parameters, 256k context, text/image/video input and text output, vLLM serving, NVIDIA Blackwell compatibility, ModelOpt v0.41.0, and NVFP4 conversion of MoE transformer-block linear operators." }, { "label": "NVIDIA Kimi K2.5 NVFP4 config", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, ModelOpt NVFP4 group_size 16, quantized Linear targets, self-attention/lm_head/vision/projector exclusions, and an FP8 kv_cache_scheme." }, { "label": "NVIDIA Kimi K2.5 NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "fixed_weight_gb" ], "notes": "hf_quant_config records producer modelopt 0.41.0, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes language_model.lm_head, all language_model.model.layers.*.self_attn modules, mm_projector, and vision_tower from weight quantization." }, { "label": "NVIDIA Kimi K2.5 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/resolve/0fd0a5e6879298d3476e3b61852a79792a35ae3d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_parameters 518016977904 and total_size 590779912064 bytes across 119 shards and 278341 tensors. Direct range-read shard headers match the index tensor count and sum exactly to 590.779912064 GB: U8 508.862398464 GB, F8_E4M3 63.607799808 GB, BF16 18.30915888 GB, and F32 0.000554912 GB. Language tensors excluding input embeddings total 587.488813984 GB; vision_tower contributes 0.833732064 GB, mm_projector 0.108555776 GB, and input embedding 2.34881024 GB, so resident-only auxiliary tensors sum to 3.29109808 GB. Routed expert tensors total 570.76144128 GB, exactly 1.48635792 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 16.727372704 GB." }, { "label": "NVIDIA Kimi K2.5 NVFP4 custom DeepSeek text runtime", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "NVIDIA Kimi K2.5 NVFP4 conditional-generation wrapper", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "auxiliary_resident_scope", "multimodal" ], "notes": "Manual review found the wrapper initializes vision_tower, mm_projector, and language_model. For ordinary text decode without pixel_values, language_model embeddings and decoder drive token generation while vision_tower and mm_projector remain resident but are not swept per generated text token." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the NVIDIA model card, served config, hf_quant_config, range-read safetensors shard headers, and manual review of the shipped custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the NVIDIA ModelOpt NVFP4 artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "nvidia--kimi-k2-6-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Kimi-K2.6-NVFP4", "title": "NVIDIA Kimi K2.6 NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 Kimi K2.6 serving artifact.", "model_family": "kimi-k2-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-K2.6", "relation": "quantized", "source": "Hugging Face model card base_model metadata, NVIDIA model card, served config, hf_quant_config, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The profile embeds the Kimi K2.6 architecture and the NVIDIA ModelOpt NVFP4/FP8-KV serving representation." }, "architecture": { "canonical_architecture_id": "kimi-k2-6", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 595.148146656, "main_resident_weight_gb": 591.857048576, "auxiliary_resident_weight_gb": 3.29109808, "fixed_weight_gb": 21.095607296, "routed_expert_weight_gb": 1.48635792, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32", "traffic_scope": "ordinary text decode through language_model layers 0-60 and lm_head, excluding full input embedding lookup and multimodal vision/projector tensors", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_tower, and mm_projector are resident for the multimodal package but not swept for each ordinary generated text token", "shared_expert_notes": "The config records one shared expert. ModelOpt quantizes Linear targets but excludes self-attention, shared_experts, lm_head, vision_tower, and mm_projector; shared expert tensors are included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the artifact mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and small F32 tensors. Routed expert tensor groups are uniform across 384 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 128 plus qk_rope_head_dim 64 across all text decoder layers." }, { "kind": "full_context", "layers": 61, "kv_heads": 64, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 128 across all text decoder layers." } ], "notes": "The custom DeepSeek/Kimi implementation computes compressed_kv internally but constructs expanded key_states and value_states before past_key_values.update. The NVIDIA hf_quant_config records FP8 KV cache, so Bounds Engine v1 charges expanded FP8 K/V cache streams for this serving artifact." }, "notes": "KimiK25ForConditionalGeneration wraps a DeepseekV3ForCausalLM language model with MoonViT vision and projector modules. This profile models ordinary text decode after optional multimodal prefill; image/video encoder and projector throughput are outside this v1 bound." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8_expanded_key_value_cache", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_expanded_key_value_cache", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored packed NVFP4 payloads, F8 scale tensors, BF16/F32 fixed tensors, and FP8 expanded K/V cache bytes. Dequantization, activation traffic, vision/video prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The served config records ModelOpt NVFP4 weights/activations with group_size 16 and the hf_quant_config records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Kimi K2.6 NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Kimi-K2.6-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving" ], "notes": "At commit 2fd3a800dedd098b8327eb49e93ebc75f85da19f, the API reports a public text-generation repo with license other, Model Optimizer library, ModelOpt/NVFP4 tags, custom_code, base_model moonshotai/Kimi-K2.6, region:us, current downloads 768850, usedStorage 595218910469 bytes, and a safetensors index with total_parameters 519536364528 and total_size 595148146656 bytes." }, { "label": "NVIDIA Kimi K2.6 NVFP4 model card", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "model_family", "max_context_tokens" ], "notes": "The card identifies the repo as a Model Optimizer quantization of Moonshot AI Kimi K2.6, an autoregressive language model with text/image/video input and text output, vLLM serving, NVIDIA Blackwell compatibility, and 256k context." }, { "label": "NVIDIA Kimi K2.6 NVFP4 config", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM text_config, dtype bfloat16, 61 hidden layers, first_k_dense_replace 1, hidden_size 7168, intermediate_size 18432, max_position_embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, ModelOpt NVFP4 group_size 16, quantized Linear targets, and exclusions for layer 0, self-attention, shared experts, lm_head, vision_tower, and mm_projector." }, { "label": "NVIDIA Kimi K2.6 NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "fixed_weight_gb" ], "notes": "hf_quant_config records producer modelopt 0.0.1.dev752, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes language_model.lm_head, language_model.model.layers.0*, all language_model.model.layers.*.self_attn modules, all shared_experts modules, mm_projector, and vision_tower from weight quantization." }, { "label": "NVIDIA Kimi K2.6 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/resolve/2fd3a800dedd098b8327eb49e93ebc75f85da19f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 595148146656 bytes across 60 shards and 277670 tensors. Direct range-read shard headers match the index total and sum exactly to 595.148146656 GB: U8 507.34301184 GB, F8_E4M3 63.41787648 GB, BF16 24.386705376 GB, and F32 0.00055296 GB. Language tensors excluding input embeddings total 591.857048576 GB; vision_tower contributes 0.833732064 GB, mm_projector 0.108555776 GB, and input embedding 2.34881024 GB, so resident-only auxiliary tensors sum to 3.29109808 GB. Routed expert tensors total 570.76144128 GB, exactly 1.48635792 GB per expert index across 384 experts. Ordinary fixed decode traffic excluding input embedding is 21.095607296 GB." }, { "label": "NVIDIA Kimi K2.6 NVFP4 custom DeepSeek text runtime", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found DeepseekV3Attention computes compressed_kv but constructs expanded key_states with 64 heads and 192 dimensions plus value_states with 64 heads and 128 dimensions before past_key_value.update. The ordinary text decoder stack uses 61 layers." }, { "label": "NVIDIA Kimi K2.6 NVFP4 conditional-generation wrapper", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/modeling_kimi_k25.py", "source_type": "manual_review", "supports": [ "auxiliary_resident_scope", "multimodal" ], "notes": "Manual review found the wrapper initializes vision_tower, mm_projector, and language_model. For ordinary text decode without pixel_values, language_model embeddings and decoder drive token generation while vision_tower and mm_projector remain resident but are not swept per generated text token." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the NVIDIA model card, served config, hf_quant_config, range-read safetensors shard headers, and manual review of the shipped custom Kimi/DeepSeek runtime code." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is an ordinary text-decode profile for the NVIDIA ModelOpt NVFP4 artifact, not a claim about optimized runtime-specific MLA cache compression." }, { "id": "nvidia--kvzap-mlp-llama-3-1-8b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/KVzap-mlp-Llama-3.1-8B-Instruct", "title": "NVIDIA KVzap MLP Llama 3.1 8B Instruct F32", "summary": "Unsupported fail-closed profile with exact resident tensor evidence for NVIDIA's F32 KVzap MLP pruning surrogate for Llama 3.1 8B Instruct.", "model_family": "kvzap-kv-pruning-sidecar-mlp", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B-Instruct", "relation": "derived_package", "source": "Repository name, model card, overview, and config", "config_compatible": false, "notes": "The checkpoint is trained to consume hidden states from a target Llama 3.1 8B Instruct model and predict per-token KV-head pruning scores. It is not a finetuned copy of the host LLM and does not expose the host model's decoder layers, vocabulary, or token-generation loop." }, "architecture": { "canonical_architecture_id": "kvzap-mlp-pruning-surrogate", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 0.067256576, "parameter_scope": "safetensors_header_f32_kvzap_sidecar", "notes": "The F32 safetensors header records 128 tensors under layers.*, totaling 67,256,576 parameters and 0.269026304 GB of tensor payload. The config records input_dim 4096, hidden_dim 512, output_dim 8, and n_modules 32, matching a two-layer MLP per target transformer module." }, "kv_adapter": { "kind": "unknown", "reason": "KVzap is not a standalone autoregressive text model. It maps host-transformer hidden states to KV-pruning scores, and the actual decode traffic depends on a separate host LLM, pruning threshold, kept-token policy, local window, KVpress integration, and downstream quality target.", "notes": "A future production adapter would need to model paired host+pruner execution: host model weight traffic, host KV allocation/read traffic after pruning, KVzap sidecar compute/traffic per layer and token, threshold policy, compression ratio, and quality constraints." }, "notes": "This profile intentionally fails closed even though the sidecar tensor bytes are accessible, because Bounds Engine v1 compares standalone target-model ordinary text-decode bounds." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-kv-pruning-sidecar-mlp", "dequantization_notes": "No quantized representation is assumed; the checkpoint stores F32 sidecar weights. Production target-model tok/s remains disabled because this is a KV-pruning auxiliary model, not the host LLM.", "notes": "The README usage path imports kvpress and applies KVzapPress inside a custom KVPressTextGenerationPipeline around a separate base model." }, "evidence": [ { "label": "NVIDIA KVzap MLP Llama 3.1 8B HF API metadata", "url": "https://huggingface.co/api/models/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b" ], "notes": "The current HF CLI/API response records commit fb936ee2cfebac42c82453100145ff17083c7552, public Apache-2.0 access, pipeline_tag other, library_name transformers, KVzapModel config, 140859 downloads, region:us, and safetensors parameters F32 67256576, total 67256576." }, { "label": "NVIDIA KVzap MLP Llama 3.1 8B config", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct/raw/fb936ee2cfebac42c82453100145ff17083c7552/config.json", "source_type": "config", "supports": [ "model_family", "serving", "unsupported_reason" ], "notes": "The pinned config records architectures KVzapModel, model_type kvzap, dtype float32, input_dim 4096, hidden_dim 512, output_dim 8, n_modules 32, and no text-token context, attention, vocabulary, or KV-cache geometry for standalone generation." }, { "label": "NVIDIA KVzap model card and overview", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "unsupported_reason", "runtime_format" ], "notes": "The README states KVzap is a KV cache pruning method used through kvpress with a separate text-generation model and predicts importance scores for every KV pair. The overview says the model consumes text-derived hidden-state tensors shaped (T, D_h), outputs numeric scores shaped (T, H), does not take raw text directly, and is a feed-forward MLP/linear surrogate rather than a transformer decoder." }, { "label": "NVIDIA KVzap MLP Llama 3.1 8B safetensors header audit", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct/resolve/fb936ee2cfebac42c82453100145ff17083c7552/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "serving", "unsupported_reason" ], "notes": "A direct range read found Content-Length 269037488 bytes, an 11176-byte safetensors header, and 269026304 tensor bytes across 128 F32 tensors under layers.*. The tensors total 67,256,576 F32 parameters. Example per-module tensors have shapes [512, 4096], [512], [8, 512], and [8], matching the config's MLP geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from current HF CLI/API metadata, pinned config, README, overview, and direct safetensors header range read. Marked unsupported because Bounds Engine v1 lacks a paired host-LLM plus KV-pruning-sidecar adapter." }, "unsupported_reason": "KVzap MLP checkpoints are auxiliary KV-pruning sidecars that score hidden states from a separate host LLM. Bounds Engine v1 does not model paired host+pruner execution, pruning thresholds, compression ratios, kept-token policies, or quality constraints, so standalone tok/s bounds would be misleading.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports explicit paired host-model and KV-pruning-sidecar profiles." }, { "id": "nvidia--kvzap-mlp-qwen3-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/KVzap-mlp-Qwen3-8B", "title": "NVIDIA KVzap MLP Qwen3 8B F32", "summary": "Unsupported profile stub for the F32 KVzap MLP auxiliary KV-cache pruning model for Qwen3 8B.", "model_family": "kvzap-auxiliary-pruner", "architecture": { "canonical_architecture_id": "kvzap-mlp-qwen3-8b", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.075663648, "swept_params_b": 0.075663648, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.302654592, "swept_weight_gb": 0.302654592, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "direct_safetensors_header_f32", "swept_parameter_scope": "recorded auxiliary scorer tensor payload", "notes": "The Hugging Face API and direct safetensors header both record 75663648 F32 parameters. This is the auxiliary KVzap pruning scorer, not the Qwen3-8B language model weights." }, "kv_adapter": { "kind": "unknown", "reason": "KVzap is an auxiliary model that scores hidden states to prune another model's KV cache through kvpress/DMS. It is not a decoder-only language model and has no self-contained token-generation KV cache or LM output projection to model with Bounds Engine v1.", "notes": "Do not infer Qwen3-8B decode traffic from this repo name. The model card usage loads Qwen/Qwen3-8B separately and instantiates KVzapPress(model_type=\"mlp\") as a compression module." }, "notes": "The config records KVzapModel with input_dim 4096, hidden_dim 512, output_dim 8, n_modules 36, and dtype float32. Bounds Engine v1 has no adapter for auxiliary KV pruning overhead or the changed downstream LM KV read pattern." }, "serving": { "weight_format": "fp32", "weight_bytes_per_param": 4, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-kvzap-auxiliary-pruner", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because the repo is an auxiliary pruning scorer rather than an LM.", "notes": "F32 weight dtype comes from the Hugging Face API safetensors metadata, served config, and direct safetensors header." }, "evidence": [ { "label": "NVIDIA KVzap MLP Qwen3 8B API metadata", "url": "https://huggingface.co/api/models/nvidia/KVzap-mlp-Qwen3-8B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "serving", "unsupported_reason" ], "notes": "At repo SHA bd5c5917846617da4311539859c137a262a6348b, the API records a public Apache-2.0 Transformers safetensors repo with pipeline other, kvzap, nvidia, pytorch, dataset nvidia/Nemotron-Pretraining-Dataset-sample, endpoints_compatible, and region:us tags. Current downloads are 693791. The API safetensors block reports F32: 75663648 and total: 75663648." }, { "label": "NVIDIA KVzap MLP Qwen3 8B model card", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B/blob/bd5c5917846617da4311539859c137a262a6348b/README.md", "source_type": "model_card", "supports": [ "architecture", "unsupported_reason" ], "notes": "The card describes KVzap as a lightweight model applied to hidden states to predict importance scores for every KV pair and prune scores below a threshold. The usage example separately loads model = Qwen/Qwen3-8B in a kv-press-text-generation pipeline and then creates DMSPress(KVzapPress(model_type=\"mlp\"), threshold=-4)." }, { "label": "NVIDIA KVzap MLP Qwen3 8B config", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B/raw/bd5c5917846617da4311539859c137a262a6348b/config.json", "source_type": "config", "supports": [ "architecture", "serving", "unsupported_reason" ], "notes": "The config records architectures KVzapModel, model_type kvzap, dtype float32, input_dim 4096, hidden_dim 512, output_dim 8, n_modules 36, and transformers_version 4.57.3. It does not describe a language-model decoder, attention heads, KV heads, or LM context window." }, { "label": "NVIDIA KVzap MLP Qwen3 8B safetensors header", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B/resolve/bd5c5917846617da4311539859c137a262a6348b/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "serving", "resident_weight_gb" ], "notes": "A direct range-read of the single safetensors header found 144 F32 tensors, header length 12592 bytes, and tensor payload 302654592 bytes / 0.302654592 GB. The linked-object HEAD check resolved to the pinned commit with linked size 302667192 bytes, leaving 12600 bytes of safetensors header/container overhead outside tensor payloads." } ], "unsupported_reason": "KVzap MLP is an auxiliary hidden-state scorer for pruning another model's KV cache. It is not a standalone decoder-only language model, and Bounds Engine v1 has no adapter for auxiliary pruning overhead or the altered downstream LM KV traffic.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports composite LM plus KV-pruner workloads." }, { "id": "nvidia--llama-3-1-8b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3.1-8B-Instruct-FP8", "title": "NVIDIA Llama 3.1 8B Instruct FP8", "summary": "Audited memory-side bounds profile for NVIDIA's ModelOpt FP8 package of Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, NVIDIA model card, public served config, ModelOpt quantization config, and safetensors index/header range reads", "config_compatible": true, "notes": "The API and model card identify meta-llama/Llama-3.1-8B-Instruct as the base model, while the card describes this repo as the quantized version produced with NVIDIA TensorRT Model Optimizer. The base Meta repo is gated in this audit environment, but the public NVIDIA config directly records the LlamaForCausalLM geometry used by this profile." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.03026176, "swept_params_b": 7.504925184, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 9.081202688, "swept_weight_gb": 8.030529536, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "safetensors_index_total_size_and_range_read_stored_fp8_bf16_f32_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings false, and the safetensors index stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors. The input embedding is resident-only for ordinary decode; the separate lm_head output projection remains in swept decode traffic. This ModelOpt FP8 artifact also stores per-tensor input_scale, weight_scale, k_scale, and v_scale F32 tensors for quantized modules, and those tiny scale tensors are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The public config records 32 layers, 8 KV heads, 128 head dimension, 131072 max position embeddings, and Llama 3 RoPE scaling. The NVIDIA ModelOpt quantization config records kv_cache_quant_algo FP8, so this profile charges one byte per KV scalar." }, "notes": "Dense LlamaForCausalLM profile using the served NVIDIA ModelOpt FP8 config and exact stored safetensors bytes." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-or-tensorrt-llm-modelopt-fp8-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus BF16 embeddings, lm_head, norms, and F32 ModelOpt scale tensors from safetensors headers. FP8 KV is charged from hf_quant_config.json. Activation quantization, FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16. The ModelOpt quantization config records FP8 weights and FP8 KV cache, with lm_head excluded from weight quantization." }, "evidence": [ { "label": "NVIDIA Llama 3.1 8B Instruct FP8 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3.1-8B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format" ], "notes": "At commit 42d9515ebd69eea3a87351d079c671c3c5ff0a31, the live API records a public non-gated text-generation repo with transformers, safetensors, Llama 3.1 license metadata, base_model meta-llama/Llama-3.1-8B-Instruct, endpoints_compatible, region:us, and 240873 downloads. The API safetensors block reports BF16: 1050939392, F8_E4M3: 6979321856, total: 8030261248; the API total excludes the 512 F32 scale scalars found in the direct headers." }, { "label": "NVIDIA Llama 3.1 8B Instruct FP8 model card", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "quantized_module_scope", "max_context_tokens" ], "notes": "The card describes this as the FP8 quantized version of Meta Llama 3.1 8B Instruct produced with NVIDIA TensorRT Model Optimizer v0.27.0. It states context length up to 128K, TensorRT-LLM and vLLM support, FP8 weight and activation quantization, only Linear operators within transformer blocks quantized, H100 deployment, and about 1.3x speedup in NVIDIA's throughput benchmark." }, { "label": "NVIDIA Llama 3.1 8B Instruct FP8 config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, bfloat16, hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, explicit head_dim 128, 131072 max position embeddings, rope_theta 500000, Llama 3 RoPE scaling from original 8192 by factor 8, vocab size 128256, tie_word_embeddings false, and use_cache true." }, { "label": "NVIDIA Llama 3.1 8B Instruct FP8 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_adapter", "serving", "quantized_module_scope" ], "notes": "The ModelOpt config records producer version 0.27.0, quant_algo FP8, kv_cache_quant_algo FP8, and exclude_modules [lm_head]." }, { "label": "NVIDIA Llama 3.1 8B Instruct FP8 safetensors index and headers", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 9081202688 bytes across two safetensors shards. Range-reading both shard headers found 803 tensors totaling exactly 9.081202688 GB: 6.979321856 GB F8_E4M3 tensors, 2.101878784 GB BF16 tensors, and 0.000002048 GB F32 scale tensors. model.embed_tokens.weight is BF16 with shape [128256, 4096] and contributes 525336576 tensor elements / 1.050673152 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, model.norm.weight, lm_head.weight, 224 input_scale tensors, 224 weight_scale tensors, 32 k_scale tensors, and 32 v_scale tensors total 8.030529536 GB swept traffic. Linked-object HEAD checks resolved both shards to 9081287040 total bytes, leaving 84352 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "RedHatAI Llama 3.1 8B FP8 config comparison", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison against the already audited RedHatAI static FP8 Llama 3.1 8B Instruct config found matching architecture, model type, dtype, hidden size, intermediate size, layer count, attention heads, KV heads, max-position embeddings, RoPE scaling, RoPE theta, vocabulary size, tied-embedding setting, and cache setting. The NVIDIA config explicitly records head_dim 128 and carries ModelOpt FP8 KV metadata in hf_quant_config.json instead of RedHatAI's compressed-tensors config." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current NVIDIA API metadata, model card, served ModelOpt FP8 config, hf_quant_config.json, safetensors index, linked-object HEAD checks, direct shard header range reads, and comparison against the existing RedHatAI Llama 3.1 8B FP8 profile." }, "notes": "This profile supersedes the scraped metadata estimate. Unlike the RedHatAI static FP8 package, the NVIDIA ModelOpt quantization config records FP8 KV cache, so production bounds use half the default BF16 KV traffic for this repo." }, { "id": "nvidia--llama-3-1-8b-instruct-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3.1-8B-Instruct-NVFP4", "title": "NVIDIA Llama 3.1 8B Instruct NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 package of Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, NVIDIA model card, public served config, ModelOpt quantization config, safetensors index/header range reads, and comparison against the audited NVIDIA FP8 sibling", "config_compatible": true, "notes": "The API and model card identify meta-llama/Llama-3.1-8B-Instruct as the base model. The base Meta repo is gated in this audit environment, but this public NVIDIA artifact directly records the LlamaForCausalLM geometry, and a field-by-field comparison found the same architecture fields as the already audited nvidia/Llama-3.1-8B-Instruct-FP8 profile." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.54060032, "swept_params_b": 4.015263744, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 6.027749376, "swept_weight_gb": 4.977076224, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "safetensors_index_total_parameters_bf16_u8_storage_accounting_plus_header_read_f8_and_f32_side_tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, lm_head.weight output projection, and ModelOpt scale tensors", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The config records tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors. The API/index total_parameters count BF16 plus U8 storage-accounting tensor elements; F8_E4M3 and tiny F32 side tensors are charged in byte traffic from direct safetensors headers." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The public config records 32 layers, 8 KV heads, 128 head dimension, 131072 max position embeddings, and Llama 3 RoPE scaling. The NVIDIA ModelOpt quantization config records kv_cache_quant_algo FP8 and config.json records a static 8-bit float KV cache scheme, so this profile charges one byte per KV scalar." }, "notes": "Dense LlamaForCausalLM profile using the served NVIDIA ModelOpt NVFP4 config and exact stored safetensors bytes." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "tensorrt-llm-modelopt-nvfp4-fp8-kv-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored tensor bytes, including ModelOpt NVFP4 U8 payload tensors, F8_E4M3 scale tensors, BF16 embeddings/lm_head/norms, tiny F32 scale tensors, and FP8 KV bytes. Dequantization, activation traffic, TensorRT-LLM/vLLM scheduling, and compute overhead are outside Bounds Engine v1.", "notes": "config.json records quant_method modelopt, quant_algo NVFP4, group_size 16, lm_head ignored, and a static 8-bit float KV cache scheme. hf_quant_config.json independently records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Llama 3.1 8B Instruct NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3.1-8B-Instruct-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "weight_format", "total_params_b", "commit_sha" ], "notes": "At commit bdb54e24298451af785c0ac63c1b485e9b7400a2, the API records a public non-gated Model Optimizer repo with safetensors, llama, nvidia, ModelOpt, quantized, FP4/fp4, base_model:meta-llama/Llama-3.1-8B-Instruct, license:other, 8-bit, modelopt, region:us, and 280705 downloads. The API safetensors block records BF16 1050939392, F8_E4M3 436207616, U8 3489660928, and total 4540600320 storage-accounting tensor elements." }, { "label": "NVIDIA Llama 3.1 8B Instruct NVFP4 model card", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "hardware", "max_context_tokens" ], "notes": "The card describes the artifact as an FP4 quantized language model of Meta's Llama 3.1 8B Instruct, states context length up to 128000, lists TensorRT-LLM, vLLM, and SGLang support, targets NVIDIA Blackwell, and says only linear operators within transformer blocks are quantized." }, { "label": "NVIDIA Llama 3.1 8B Instruct NVFP4 config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records LlamaForCausalLM, bfloat16, hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, rope_theta 500000, Llama 3 RoPE scaling from original 8192 by factor 8, vocab size 128256, tie_word_embeddings false, use_cache true, and ModelOpt NVFP4 quantization with group_size 16, lm_head ignored, and static 8-bit float KV cache." }, { "label": "NVIDIA Llama 3.1 8B Instruct NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt 0.37.0.dev5+g76fb12d47.d20250905, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and lm_head exclusion." }, { "label": "NVIDIA Llama 3.1 8B Instruct NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "dtype_split" ], "notes": "The index records total_parameters 4540600320 and total_size 6027749376 bytes across two shards. Range-read safetensors headers found 1027 tensors totaling exactly 6.027749376 GB: BF16 2.101878784 GB, U8 3.489660928 GB, F8_E4M3 0.436207616 GB, and F32 0.000002048 GB. model.embed_tokens.weight has shape [128256, 4096] and contributes 0.525336576B parameters / 1.050673152 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 4.977076224 GB. Linked-object HEAD checks resolved both shards to 6.027861376 GB, leaving 112000 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "NVIDIA Llama 3.1 8B FP8 config comparison", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison against the already audited NVIDIA ModelOpt FP8 Llama 3.1 8B Instruct config found matching architecture, model type, dtype, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, max-position embeddings, RoPE scaling, RoPE theta, vocabulary size, tied-embedding setting, attention/mlp bias settings, and cache setting. The NVFP4 package differs in its ModelOpt quantization block and exact stored tensor representation." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned NVIDIA model card, served ModelOpt NVFP4 config, hf_quant_config.json, generation config, safetensors index, direct safetensors shard header range reads, linked-object HEAD checks, and field-by-field comparison against the audited NVIDIA Llama 3.1 8B FP8 sibling config." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged BF16 KV instead of the artifact's declared FP8 KV." }, { "id": "nvidia--llama-3-1-nemotron-nano-vl-8b-v1", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", "title": "NVIDIA Llama 3.1 Nemotron Nano VL 8B V1 BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Llama 3.1 Nemotron Nano VL 8B V1 vision-language repo.", "model_family": "llama-3-1-nemotron-nano-vl-dense", "architecture": { "canonical_architecture_id": "llama-3-1-nemotron-nano-vl-8b", "max_context_tokens": 16384, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.721768454, "swept_params_b": 7.505973248, "auxiliary_resident_params_b": 1.215795206, "resident_weight_gb": 17.443536908, "swept_weight_gb": 15.011946496, "auxiliary_resident_weight_gb": 2.431590412, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode through language_model.model.layers.*, language_model.model.norm.*, and language_model.lm_head.weight", "auxiliary_scope": "language_model.model.embed_tokens.weight, vision_model tensors, and mlp1 projector tensors are resident for the multimodal package but are not swept as full matrices for each generated text token", "notes": "The runtime wraps a BF16 LlamaForCausalLM language_model, a C-RADIO vision_model, and an mlp1 vision projector. The config records tie_word_embeddings false and the safetensors header stores a separate language_model.lm_head.weight, so lm_head.weight is swept for ordinary text decode while language_model.model.embed_tokens.weight is resident-only for token lookup." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The embedded llm_config records 32 Llama decoder layers, hidden size 4096, 32 query heads, and 8 key/value heads. The head dimension is 4096 / 32 = 128, matching k_proj and v_proj safetensors shapes [1024, 4096]. The top-level config and model card limit the supported input plus output token budget to 16K even though the embedded Llama RoPE config records 131072 max position embeddings." }, "notes": "Llama_Nemotron_Nano_VL is a composed VLM with a Llama-3.1-8B-Instruct language encoder, C-RADIOv2-H vision encoder, and BF16 MLP projector. This profile models ordinary generated text-token decode after any image prefill. Vision encoder, image tiling, projector prefill, and image-token replacement work are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-llama-vlm-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Vision encoder execution, projector prefill, image-token embedding replacement, activation traffic, kernels, and scheduler behavior are outside Bounds Engine v1.", "notes": "The top-level config and embedded llm_config record bfloat16, and the safetensors header records only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "NVIDIA Llama 3.1 Nemotron Nano VL model card", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "max_context_tokens", "multimodal" ], "notes": "The model card records NVIDIA Open Model License terms, image-text-to-text packaging, a C-RADIOv2-H vision encoder, Llama-3.1-8B-Instruct language encoder, single-image inference support, input plus output token budget of 16K, and TensorRT-LLM/H100 deployment notes." }, { "label": "NVIDIA Llama 3.1 Nemotron Nano VL API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", "source_type": "derived_calculation", "supports": [ "repo", "revision", "downloads", "weight_format", "total_params_b", "pipeline" ], "notes": "At repo SHA 437f4e28b989cc2d9a16b6767cc930cdf48797ff, the API records a public transformers image-text-to-text repo with current downloads 1220336, region:us tag, and safetensors parameters BF16: 8721768454." }, { "label": "NVIDIA Llama 3.1 Nemotron Nano VL resolved config", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "vision_projector", "tie_word_embeddings" ], "notes": "The resolved config records Llama_Nemotron_Nano_VL, torch_dtype bfloat16, max_sequence_length 16384, force_image_size 512, downsample_ratio 0.5, projector_hidden_size 4096, vit_hidden_size 1280, and an embedded Llama llm_config with 32 hidden layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, vocab size 128512, tie_word_embeddings false, Llama 3 RoPE scaling to 131072, and BF16 dtype. Raw config.json is stored through Git LFS, so this audit uses the resolved pinned file." }, { "label": "NVIDIA Llama 3.1 Nemotron Nano VL custom runtime code", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/modeling.py", "source_type": "manual_review", "supports": [ "weight_adapter", "kv_adapter", "vision_projector", "ordinary_text_decode_scope" ], "notes": "Manual review found Llama_Nemotron_Nano_VL initializes language_model = AutoModelForCausalLM.from_config(config.llm_config), vision_model = AutoModel.from_config(config.vision_config), and mlp1 as a LayerNorm plus two Linear projector stack. Forward/generate replace image placeholder token embeddings with projected vision features, then call language_model.generate with use_cache=True. With no new image processing during decode, ordinary generated text-token traffic is the Llama language model plus untied lm_head." }, { "label": "NVIDIA Llama 3.1 Nemotron Nano VL image processor", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/image_processing.py", "source_type": "manual_review", "supports": [ "vision_prefill_scope" ], "notes": "Manual review found the image processor dynamically tiles images up to max_num_tiles 12 and returns pixel_values and num_patches for prefill. Image tiling and vision feature extraction are resident/prefill work, not per-generated-token text decode traffic." }, { "label": "NVIDIA Llama 3.1 Nemotron Nano VL safetensors header", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_resident_scope" ], "notes": "The single safetensors file was range-read directly. The file size is 17.443626956 GB, with a 90040-byte header and 17.443536908 GB of BF16 tensor payloads across 686 tensors, matching 8721768454 parameters. language_model tensors total 16.064716800 GB, vision_model tensors total 1.303285772 GB, and mlp1 tensors total 0.075534336 GB. The swept ordinary text subset is language_model.model.layers.* plus language_model.model.norm.* plus language_model.lm_head.weight: 7.505973248B parameters / 15.011946496 GB. Resident-only tensors are language_model.model.embed_tokens.weight plus vision_model and mlp1: 1.215795206B parameters / 2.431590412 GB. The separate language_model.model.embed_tokens.weight and language_model.lm_head.weight tensors each have shape [128512, 4096] and 1.052770304 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, model card, resolved pinned config, custom runtime code, image processor code, HEAD/content-length checks, and direct safetensors header range reads." }, "notes": "This self-contained profile deliberately uses the top-level VLM 16K token budget for max_context_tokens while preserving the embedded Llama 131K RoPE detail in evidence notes. The memory-side throughput bound is for ordinary text decode after multimodal prefill." }, { "id": "nvidia--llama-3-3-nemotron-super-49b-v1-5-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8", "title": "Llama 3.3 Nemotron Super 49B v1.5 FP8", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's ModelOpt FP8 package of Llama 3.3 Nemotron Super 49B v1.5.", "model_family": "llama-3.3-nemotron-nas-dense", "base_model_proof": { "base_model": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5", "relation": "quantized", "source": "Hugging Face model card, served config comparison with the audited BF16 sibling, ModelOpt quantization config, and safetensors headers", "config_compatible": true, "notes": "The FP8 model card describes the same Llama 3.3 Nemotron Super 49B v1.5 NAS architecture family and the served config is byte-identical to the audited BF16 v1.5 config for every checked architecture field, including 80 block configs, 49 active attention blocks, 31 attention no-op blocks, FFN multipliers, untied embeddings, and 131072 max positions. The FP8 repo adds ModelOpt serving metadata and FP8/BF16/F32 stored tensor dtypes." }, "architecture": { "canonical_architecture_id": "llama-3.3-nemotron-super-49b-v1.5", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 49.867146186, "swept_params_b": 48.816473034, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 51.96956036, "swept_weight_gb": 49.868214056, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "safetensors_header_modelopt_fp8_bf16_f32_tensor_elements", "swept_parameter_scope": "ordinary text decode through model.layers, model.norm, lm_head.weight, and ModelOpt scale tensors, excluding model.embed_tokens.weight after token embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "Range-read safetensors headers record exact mixed ModelOpt storage rather than an ideal one-byte-per-parameter checkpoint. The config records tie_word_embeddings false, and the checkpoint stores separate BF16 model.embed_tokens.weight and lm_head.weight tensors. The input embedding is resident-only for ordinary decode; the separate lm_head output projection remains in swept decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 49, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served block configs contain 49 active attention blocks and 31 no-op attention blocks. Active attention blocks have window_length null and n_heads_in_group 8. The audited DeciLM runtime derives num_key_value_heads as num_attention_heads / n_heads_in_group, so 64 query heads produce 8 stored KV heads. The ModelOpt quantization config records kv_cache_quant_algo FP8, so this profile charges one byte per KV scalar." }, "notes": "The DeciLM/Nemotron-NAS config records 80 dense transformer blocks with no MoE routing. Attention is skipped in 31 blocks, while all 80 blocks retain FFNs with ffn_mult distribution {0.5: 6, 1: 8, 1.3125: 10, 2.625: 6, 3.28125: 1, 5.25: 49}." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-fp8-decilm-nemotron-nas-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, including F8_E4M3 weights, BF16 embeddings/head/norm tensors, F32 ModelOpt scale tensors, and FP8 KV cache bytes from hf_quant_config.json. FP8 dequantization, activation traffic, tensor-parallel communication, scheduler overhead, and compute are outside Bounds Engine v1.", "notes": "The served config records torch_dtype bfloat16. The ModelOpt quantization sidecar records quant_algo FP8, kv_cache_quant_algo FP8, and exclude_modules [lm_head]. The model card's vLLM examples use --quantization=modelopt." }, "evidence": [ { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "weight_format" ], "notes": "At repo SHA 04822723e77e036ddf2d24e83c6d469d3b009252, the HF API records a public non-gated transformers text-generation repo with NVIDIA Open Model License metadata, custom_code, nemotron-nas, llama-3, region:us, 233019 downloads, and safetensors parameters BF16 2102411264, F8_E4M3 47764733952, total 49867145216." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "model_family", "max_context_tokens", "serving" ], "notes": "The model card states that Llama-3.3-Nemotron-Super-49B-v1.5-FP8 is a derivative of Meta Llama-3.3-70B-Instruct, customized through Neural Architecture Search, with skip-attention blocks, variable FFN expansion ratios, context length up to 131072 tokens, and vLLM serving with --quantization=modelopt." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 served config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/config.json", "source_type": "config", "supports": [ "model_family", "layers", "active_attention_layers", "kv_heads", "head_dim", "max_context_tokens", "ffn_layout", "serving" ], "notes": "The served config records model_type nemotron-nas, DeciLMForCausalLM, torch_dtype bfloat16, hidden_size 8192, 64 attention heads, 80 block_configs, max_position_embeddings 131072, Llama 3 rope_scaling factor 16, vocab_size 128256, and tie_word_embeddings false. Manual comparison found no differences in audited architecture fields versus the pinned BF16 v1.5 config. Manual block_config counting found 49 active attention blocks, 31 attention no-op blocks, no sliding-window or sink attention, active attention n_heads_in_group 8 throughout, and the same FFN multiplier distribution as the BF16 sibling." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "quantized_module_scope" ], "notes": "hf_quant_config.json records ModelOpt 0.28.1.dev55, quant_algo FP8, kv_cache_quant_algo FP8, and exclude_modules [lm_head]." }, { "label": "NVIDIA DeciLM custom runtime code from audited BF16 sibling", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/modeling_decilm.py", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_no_op", "kv_heads", "cache_behavior" ], "notes": "The FP8 repo's pinned config has the same auto_map and byte-identical architecture fields as the audited BF16 sibling, but the FP8 repo does not include modeling_decilm.py at the pinned commit. The BF16 DeciLM runtime review found DeciLMDecoderLayer skipping self_attn when attention_config.no_op is true and DeciLMAttention deriving num_key_value_heads as num_attention_heads / n_heads_in_group before storing K/V in the cache." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 block config dataclasses", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/block_config.py", "source_type": "manual_review", "supports": [ "attention_no_op", "sliding_window", "ffn_layout" ], "notes": "Manual review of the same block_config dataclass surface used by the BF16 sibling found AttentionConfig clearing n_heads_in_group/window fields for no-op attention blocks, requiring n_heads_in_group for active attention, and exposing window_length only for sliding or sink attention. The pinned FP8 config uses null window_length for every block." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 FP8 safetensors index and headers", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_parameters 49867145216 and total_size 51969560360 bytes across 11 safetensors shards. Range-reading all shard headers found 1538 tensors totaling exactly 51.969560360 GB: F8_E4M3 47.764733952 GB, BF16 4.204822528 GB, and F32 scale tensors 0.000003880 GB. Direct header tensor elements total 49.867146186B, 970 scalar F32 scale elements above the API/index logical total. model.embed_tokens.weight is BF16 [128256, 8192] and contributes 1.050673152B elements / 2.101346304 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, model.norm.weight, lm_head.weight, 436 input_scale tensors, 436 weight_scale tensors, 49 k_scale tensors, and 49 v_scale tensors total 48.816473034B elements / 49.868214056 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, audited BF16 config/runtime comparison, FP8 ModelOpt quantization config, block_config dataclasses, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "Production code must use this self-contained profile rather than deriving KV heads, active attention layers, FP8 KV cache, or exact mixed FP8/BF16/F32 weight traffic from the repo name. The generated catalog row had assumed 64 KV heads across all 80 layers and one-byte resident weights, missing both the audited 49 active attention layers and the BF16/F32 side tensors." }, { "id": "nvidia--llama-3-3-nemotron-super-49b-v1-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5", "title": "Llama 3.3 Nemotron Super 49B v1.5 BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Llama 3.3 Nemotron Super 49B v1.5 DeciLM/Nemotron-NAS repo.", "model_family": "llama-3.3-nemotron-nas-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.3-70B-Instruct", "relation": "derived_package", "source": "Hugging Face model card, served config, and custom DeciLM runtime review", "config_compatible": false, "notes": "The model card describes this repo as a derivative of Meta Llama-3.3-70B-Instruct customized through Neural Architecture Search. The served DeciLM config is not geometry-compatible with the reference model: it records 80 non-repetitive blocks, 49 active attention blocks, 31 no-op attention blocks, and variable FFN expansion ratios." }, "architecture": { "canonical_architecture_id": "llama-3.3-nemotron-super-49b-v1.5", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 49.867145216, "swept_params_b": 48.816472064, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 99.734290432, "swept_weight_gb": 97.632944128, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "all BF16 tensors in the pinned safetensors package", "swept_parameter_scope": "ordinary text decode through model.layers, model.norm, and the separately stored lm_head.weight, excluding model.embed_tokens.weight after token embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "All 21 safetensors shard headers were range-read directly. The headers match the safetensors index total_size exactly and contain only BF16 tensors." }, "kv_adapter": { "kind": "full_context", "layers": 49, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served block_configs contain 49 active attention blocks and 31 no-op attention blocks. All active attention blocks have window_length null and n_heads_in_group 8. The custom DeciLMAttention code derives num_key_value_heads as num_attention_heads / n_heads_in_group, so 64 query heads produce 8 stored KV heads before repeat_kv." }, "notes": "The DeciLM/Nemotron-NAS config records 80 dense transformer blocks with no MoE routing. Attention is skipped in 31 blocks, while all 80 blocks retain FFNs with ffn_mult distribution {0.5: 6, 1: 8, 1.3125: 10, 2.625: 6, 3.28125: 1, 5.25: 49}." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-or-vllm-bf16-decilm-nemotron-nas", "dequantization_notes": "No quantized weight or KV representation is assumed for this BF16 repo. Bounds Engine v1 charges exact BF16 safetensors bytes, BF16 K/V cache allocation, and BF16 K/V cache read traffic. Compute, activation traffic, tensor-parallel communication, and vLLM scheduling overhead are outside this memory-side bound.", "notes": "The model card gives vLLM examples with trust-remote-code, tensor_parallel_size 8, max_model_len 65536, and no explicit KV-cache quantization flag. The served config records torch_dtype bfloat16, so this profile charges BF16 KV." }, "evidence": [ { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "weight_format" ], "notes": "At repo SHA 420ba7d28211abf116b8b103ab700d92619daf98, the HF API records a public transformers text-generation repo with NVIDIA Open Model License metadata, custom_code, nemotron-nas, llama-3, region:us, 259026 downloads, and safetensors parameters BF16 49867145216." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "model_family", "max_context_tokens", "serving" ], "notes": "The model card states that Llama-3.3-Nemotron-Super-49B-v1.5 is a derivative of Meta Llama-3.3-70B-Instruct, customized through Neural Architecture Search, with skip-attention blocks, variable FFN expansion ratios, and context length up to 131072 tokens. The vLLM example uses trust-remote-code and no explicit KV quantization option." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 served config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/config.json", "source_type": "config", "supports": [ "model_family", "layers", "active_attention_layers", "kv_heads", "head_dim", "max_context_tokens", "ffn_layout", "serving" ], "notes": "The served config records model_type nemotron-nas, DeciLMForCausalLM, torch_dtype bfloat16, hidden_size 8192, 64 attention heads, 80 block_configs, max_position_embeddings 131072, Llama 3 rope_scaling factor 16, vocab_size 128256, and tie_word_embeddings false. Manual block_config counting found 49 active attention blocks, 31 attention no-op blocks, no sliding-window or sink attention, and active attention n_heads_in_group 8 throughout." }, { "label": "NVIDIA DeciLM custom runtime code", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/modeling_decilm.py", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_no_op", "kv_heads", "cache_behavior" ], "notes": "Manual review found DeciLMDecoderLayer skipping self_attn when attention_config.no_op is true, and DeciLMAttention deriving num_key_value_heads as num_attention_heads / n_heads_in_group. K and V are projected to num_key_value_heads * head_dim, stored in the cache via past_key_value.update, and only then expanded with repeat_kv for attention." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 block config dataclasses", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/block_config.py", "source_type": "manual_review", "supports": [ "attention_no_op", "sliding_window", "ffn_layout" ], "notes": "Manual review found AttentionConfig clears n_heads_in_group/window fields for no-op attention blocks, requires n_heads_in_group for active attention, and exposes window_length only for sliding or sink attention. The pinned config uses null window_length for every block." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b" ], "notes": "The safetensors index records total_size 99734290432 bytes across 21 shards and 568 tensors. Direct range reads of all shard headers matched that total exactly: 99.734290432 GB / 49.867145216B BF16 parameters. model.embed_tokens.weight has shape [128256, 8192] and contributes 1.050673152B parameters / 2.101346304 GB resident-only. model.layers plus model.norm plus the separately stored lm_head.weight contribute 48.816472064B parameters / 97.632944128 GB swept ordinary text-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, custom DeciLM runtime code, block_config dataclasses, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "Production code must use this self-contained profile rather than deriving KV heads or active attention layers from the repo name. The generated catalog row had assumed 64 KV heads across all 80 layers; the audited DeciLM code and config show only 8 stored KV heads across 49 active full-context attention layers." }, { "id": "nvidia--llama-3-3-nemotron-super-49b-v1", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Llama-3_3-Nemotron-Super-49B-v1", "title": "Llama 3.3 Nemotron Super 49B v1 BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Llama 3.3 Nemotron Super 49B v1 DeciLM/Nemotron-NAS repo.", "model_family": "llama-3.3-nemotron-nas-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.3-70B-Instruct", "relation": "derived_package", "source": "Hugging Face model card, served config, and custom DeciLM runtime review", "config_compatible": false, "notes": "The model card describes this repo as a derivative of Meta Llama-3.3-70B-Instruct customized through Neural Architecture Search. The served DeciLM config is not geometry-compatible with the reference model: it records 80 non-repetitive blocks, 49 active attention blocks, 31 no-op attention blocks, and variable FFN expansion ratios." }, "architecture": { "canonical_architecture_id": "llama-3.3-nemotron-super-49b-v1", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 49.867145216, "swept_params_b": 48.816472064, "auxiliary_resident_params_b": 1.050673152, "resident_weight_gb": 99.734290432, "swept_weight_gb": 97.632944128, "auxiliary_resident_weight_gb": 2.101346304, "resident_parameter_scope": "all BF16 tensors in the pinned safetensors package", "swept_parameter_scope": "ordinary text decode through model.layers, model.norm, and the separately stored lm_head.weight, excluding model.embed_tokens.weight after token embedding lookup", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "All 21 safetensors shard headers were range-read directly. The headers match the safetensors index total_size exactly and contain only BF16 tensors." }, "kv_adapter": { "kind": "full_context", "layers": 49, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served block_configs contain 49 active attention blocks and 31 no-op attention blocks. All active attention blocks have window_length null and n_heads_in_group 8. The custom DeciLMAttention code derives num_key_value_heads as num_attention_heads / n_heads_in_group, so 64 query heads produce 8 stored KV heads before repeat_kv." }, "notes": "The DeciLM/Nemotron-NAS config records 80 dense transformer blocks with no MoE routing. Attention is skipped in 31 blocks, while all 80 blocks retain FFNs with ffn_mult distribution {0.5: 6, 1: 8, 1.3125: 10, 2.625: 6, 3.28125: 1, 5.25: 49}." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-or-vllm-bf16-decilm-nemotron-nas", "dequantization_notes": "No quantized weight or KV representation is assumed for this BF16 repo. Bounds Engine v1 charges exact BF16 safetensors bytes, BF16 K/V cache allocation, and BF16 K/V cache read traffic. Compute, activation traffic, tensor-parallel communication, and vLLM scheduling overhead are outside this memory-side bound.", "notes": "The model card gives vLLM examples with trust-remote-code and no explicit KV-cache quantization flag. The served config records torch_dtype bfloat16, so this profile charges BF16 KV." }, "evidence": [ { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1", "source_type": "derived_calculation", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "weight_format" ], "notes": "At repo SHA 387156d8d6868c19f3472fa607aa9bfc4f662333, the HF API records a public transformers text-generation repo with NVIDIA Open Model License metadata, custom_code, nemotron-nas, llama-3, region:us, 188263 downloads, and safetensors parameters BF16 49867145216." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1 model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "model_family", "max_context_tokens", "serving" ], "notes": "The model card states that Llama-3.3-Nemotron-Super-49B-v1 is a derivative of Meta Llama-3.3-70B-Instruct, customized through Neural Architecture Search, with skip-attention blocks, variable FFN expansion ratios, support for up to 128K context, and vLLM serving with trust-remote-code." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1 served config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/config.json", "source_type": "config", "supports": [ "model_family", "layers", "active_attention_layers", "kv_heads", "head_dim", "max_context_tokens", "ffn_layout", "serving" ], "notes": "The served config records model_type nemotron-nas, DeciLMForCausalLM, torch_dtype bfloat16, hidden_size 8192, 64 attention heads, 80 block_configs, max_position_embeddings 131072, Llama 3 rope_scaling factor 16, vocab_size 128256, and tie_word_embeddings false. Manual block_config counting found 49 active attention blocks, 31 attention no-op blocks, no sliding-window or sink attention, and active attention n_heads_in_group 8 throughout." }, { "label": "NVIDIA DeciLM custom runtime code", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/modeling_decilm.py", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_no_op", "kv_heads", "cache_behavior" ], "notes": "Manual review found DeciLMDecoderLayer skipping self_attn when attention_config.no_op is true, and DeciLMAttention deriving num_key_value_heads as num_attention_heads / n_heads_in_group. K and V are projected to num_key_value_heads * head_dim, stored in the cache via past_key_value.update, and only then expanded with repeat_kv for attention." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1 block config dataclasses", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/block_config.py", "source_type": "manual_review", "supports": [ "attention_no_op", "sliding_window", "ffn_layout" ], "notes": "Manual review found AttentionConfig clears n_heads_in_group/window fields for no-op attention blocks, requires n_heads_in_group for active attention, and exposes window_length only for sliding or sink attention. The pinned config uses null window_length for every block." }, { "label": "NVIDIA Llama 3.3 Nemotron Super 49B v1 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b" ], "notes": "The safetensors index records total_size 99734290432 bytes across 21 shards and 568 tensors. Direct range reads of all shard headers matched that total exactly: 99.734290432 GB / 49.867145216B BF16 parameters. model.embed_tokens.weight has shape [128256, 8192] and contributes 1.050673152B parameters / 2.101346304 GB resident-only. model.layers plus model.norm plus the separately stored lm_head.weight contribute 48.816472064B parameters / 97.632944128 GB swept ordinary text-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, custom DeciLM runtime code, block_config dataclasses, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "Production code must use this self-contained profile rather than deriving KV heads or active attention layers from the repo name. The generated catalog row had assumed 64 KV heads across all 80 layers; the audited DeciLM code and config show only 8 stored KV heads across 49 active full-context attention layers." }, { "id": "nvidia--llama-4-scout-17b-16e-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/Llama-4-Scout-17B-16E-Instruct-FP8", "title": "NVIDIA Llama 4 Scout 17B 16E Instruct FP8", "summary": "Unsupported profile stub for the gated NVIDIA ModelOpt FP8 package of Llama 4 Scout 17B 16E Instruct.", "model_family": "llama4-scout-multimodal-moe", "base_model_proof": { "base_model": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "relation": "quantized", "source": "Hugging Face API model metadata and NVIDIA model card", "config_compatible": false, "notes": "The public API identifies this repo as a quantized package of meta-llama/Llama-4-Scout-17B-16E-Instruct, but the served config and tensor index are gated. Production compatibility with the base cannot be verified from direct config evidence in this audit environment." }, "architecture": { "canonical_architecture_id": "llama-4-scout-17b-16e-fp8", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 108.641793536, "parameter_scope": "hf_api_safetensors_total_only_gated_tensor_layout_unverified", "notes": "The Hugging Face API records 108641793536 safetensors parameters: 2945332736 BF16 parameters and 105696460800 F8_E4M3 parameters. This API total is preserved only as package metadata; resident, active, routed-expert, multimodal, and swept text-decode traffic are not audited because config, quant config, and tensor-index files are gated." }, "kv_adapter": { "kind": "unknown", "reason": "The repo is auto-gated, and raw config, generation config, hf_quant_config, README, model.safetensors.index.json, and hf download config.json are inaccessible with the configured osolmaz CLI identity.", "notes": "Do not infer Llama 4 MoE routing, multimodal components, KV heads, context length, tied embeddings, or swept decode traffic from the repo name or from the ungated API parameter total. Replace this with an audited adapter only after direct gated config and tensor-header evidence is available." }, "notes": "The NVIDIA model card describes a ModelOpt FP8 quantized multimodal MoE package with up to 1M context and TensorRT-LLM/SGLang support, but the production bounds profile intentionally fails closed because the executable config and tensor layout are not accessible." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0271104944067775, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-gated-config", "dequantization_notes": "The API and model card identify FP8 ModelOpt quantization with BF16 side tensors, but production bounds are disabled because KV geometry, MoE routing, and swept tensor traffic are unavailable.", "notes": "The effective 1.0271104944067775 bytes/parameter value is computed from API safetensors dtype counts only: BF16 parameters charged at 2 bytes and F8_E4M3 parameters charged at 1 byte. It is not used for production throughput because status is unsupported." }, "evidence": [ { "label": "NVIDIA Llama 4 Scout FP8 API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "total_params_b", "serving", "unsupported_reason" ], "notes": "At commit 3c0175a5a4a66f8e139faf93ddeb7d2741ff3560, the API reports gated: auto, library_name Model Optimizer, license other / nvidia-open-model-license, region:us, 334755 downloads, and safetensors counts BF16 2945332736, F8_E4M3 105696460800, total 108641793536." }, { "label": "NVIDIA Llama 4 Scout FP8 model card", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8", "source_type": "model_card", "supports": [ "model_family", "base_model_proof", "serving", "unsupported_reason" ], "notes": "The rendered model card describes a ModelOpt v0.33.0 FP8 quantized package of Llama 4 Scout 17B 16E Instruct, an auto-regressive multimodal MoE model with up to 1M context, TensorRT-LLM and SGLang support, cnn_dailymail calibration, and H100 TensorRT-LLM evaluation. The card does not provide enough executable tensor-layout detail to audit bounds without gated files." }, { "label": "Gated config access check", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8/raw/3c0175a5a4a66f8e139faf93ddeb7d2741ff3560/config.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Unauthenticated raw config access returned 401 restricted-access text. hf download config.json with the configured osolmaz CLI identity also failed, so the audit environment does not have accepted-gate access." }, { "label": "Gated tensor index access check", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8/raw/3c0175a5a4a66f8e139faf93ddeb7d2741ff3560/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "unsupported_reason" ], "notes": "Raw model.safetensors.index.json, generation_config.json, hf_quant_config.json, and README.md all returned 401 restricted-access responses. Direct tensor grouping, MoE expert layout, quantization exclusions, and KV/cache settings therefore cannot be audited." }, { "label": "Meta Llama 4 Scout Instruct base model card", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The base model is also gated in this audit environment, matching the existing fail-closed base profile. This NVIDIA profile does not copy base geometry or routing from the gated base." } ], "unsupported_reason": "Gated config, quant config, and tensor index files are not accessible in this audit environment, so multimodal architecture, MoE routing, KV geometry, max context, active/swept traffic, resident component split, and exact FP8/BF16 tensor grouping cannot be verified without guessing.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after accepted-gate access allows direct config, quant config, and safetensors header evidence." }, { "id": "nvidia--locateanything-3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/LocateAnything-3B", "title": "NVIDIA LocateAnything 3B BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 LocateAnything 3B vision-language grounding repo.", "model_family": "locateanything-qwen2-grounding", "base_model_proof": { "base_model": "Qwen/Qwen2.5-3B-Instruct", "relation": "finetune", "source": "Hugging Face model card base_model metadata and served text_config", "config_compatible": false, "notes": "The repo records Qwen/Qwen2.5-3B-Instruct as the language base, but the served package wraps it in LocateAnythingForConditionalGeneration with a MoonViT vision encoder, projector, custom Magi/range attention, and hybrid MTP/AR box decoding. This profile does not inherit ordinary Qwen2.5 text-decode bounds." }, "architecture": { "canonical_architecture_id": "locateanything-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.830665968, "swept_params_b": 3.400155136, "auxiliary_resident_params_b": 0.430510832, "resident_weight_gb": 7.661331936, "swept_weight_gb": 6.800310272, "auxiliary_resident_weight_gb": 0.861021664, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "language_model tensors excluding input embedding plus lm_head; recorded for evidence only because production bounds are disabled", "auxiliary_scope": "vision_model, mlp1 projector, and language_model.model.embed_tokens.weight are resident for the package but do not define ordinary token-by-token decode throughput", "notes": "The header grouping is exact: language_model tensors total 6.800310272 GB, vision_model tensors total 0.833732064 GB, mlp1 tensors total 0.027289600 GB, and language_model.model.embed_tokens.weight plus language_model.lm_head.weight are each 0.625381376 GB." }, "kv_adapter": { "kind": "unknown", "reason": "LocateAnything uses custom hybrid generation for parallel box decoding. The model card describes Parallel Box Decoding, and the custom code switches between MTP and autoregressive modes with range/Magi attention masks, per-row KV cache packing, and task-specific box token sampling. Bounds Engine v1 only models ordinary autoregressive text-token decode with full/sliding KV or fixed recurrent state adapters.", "notes": "Do not infer production throughput from the wrapped Qwen2 decoder geometry alone. A future adapter must model MTP window size, box-token sampling, hybrid AR fallback, range-attention masks, visual feature insertion, and box-level rather than text-token-level throughput." }, "notes": "This profile intentionally fails closed even though config and tensor headers are accessible, because LocateAnything's primary generation algorithm is not ordinary text-token decode." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-locateanything-hybrid-box-decoding", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The API safetensors block and range-read shard headers record only BF16 tensors. KV/state production bounds are disabled because the generation algorithm is unsupported." }, "evidence": [ { "label": "NVIDIA LocateAnything 3B API metadata", "url": "https://huggingface.co/api/models/nvidia/LocateAnything-3B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit c32291ca5e996f5a7a485845b4f57a233936bba0, the live API records a public image-text-to-text repo with NVIDIA non-commercial license metadata, custom_code, locateanything, object-detection, grounding, base_model Qwen/Qwen2.5-3B-Instruct, region:us, 1194542 downloads, and BF16 safetensors parameters 3830665968." }, { "label": "NVIDIA LocateAnything 3B config", "url": "https://huggingface.co/nvidia/LocateAnything-3B/raw/c32291ca5e996f5a7a485845b4f57a233936bba0/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "serving", "unsupported_reason" ], "notes": "The config records LocateAnythingForConditionalGeneration, model_type locateanything, _attn_implementation magi, bfloat16 dtype, a Qwen2 text_config with 36 layers, 16 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 32768, use_cache false, causal_attn false, use_sliding_window false, and a MoonViT vision_config with 27 layers. It also records mlp_connector_layers 2 and custom box/ref/coord token IDs." }, { "label": "NVIDIA LocateAnything 3B model card", "url": "https://huggingface.co/nvidia/LocateAnything-3B", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The model card describes LocateAnything as a vision-language grounding model whose core innovation is Parallel Box Decoding, predicting complete bounding-box coordinates in a single parallel step rather than standard autoregressive token-by-token decoding." }, { "label": "NVIDIA LocateAnything 3B custom modeling and generation code", "url": "https://huggingface.co/nvidia/LocateAnything-3B/raw/c32291ca5e996f5a7a485845b4f57a233936bba0/modeling_locateanything.py", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The custom generate path supports fast, slow, and hybrid generation modes; asserts use_cache true; prepares MTP windows with mask tokens; switches between MTP and AR on error_box and box_end_ar events; slices past_key_values to generated length; and samples task-specific box/reference token patterns. The bundled batch runtime additionally packs per-row KV caches and builds Magi/range-attention scheduler masks." }, { "label": "NVIDIA LocateAnything 3B safetensors index and shard headers", "url": "https://huggingface.co/nvidia/LocateAnything-3B/raw/c32291ca5e996f5a7a485845b4f57a233936bba0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format" ], "notes": "The index records total_size 7661331936 bytes across two shards. Range-read shard headers found 770 BF16 tensors totaling 3.830665968 parameters / 7.661331936 GB, matching the index. Prefix grouping found language_model 6.800310272 GB, vision_model 0.833732064 GB, and mlp1 0.027289600 GB. language_model.model.embed_tokens.weight and language_model.lm_head.weight are separate 0.625381376 GB tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from live HF API metadata, model card, served config, custom modeling/generation code, safetensors index, and direct shard header byte grouping. Marked unsupported because Bounds Engine v1 lacks a LocateAnything/PBD hybrid decode adapter." }, "unsupported_reason": "LocateAnything-3B uses custom hybrid MTP/autoregressive Parallel Box Decoding with Magi/range attention and task-specific box-token sampling. Bounds Engine v1 only supports ordinary autoregressive text-token decode adapters, so production tok/s bounds are disabled to avoid misleading comparisons.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine has a dedicated LocateAnything/parallel-box-decoding adapter." }, { "id": "nvidia--minimax-m2-7-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/MiniMax-M2.7-NVFP4", "title": "NVIDIA MiniMax M2.7 NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's ModelOpt NVFP4 MiniMax M2.7 serving artifact.", "model_family": "minimax-m2-moe-nvfp4", "base_model_proof": { "base_model": "MiniMaxAI/MiniMax-M2.7", "relation": "quantized", "source": "Hugging Face model card/API metadata, served ModelOpt config, hf_quant_config, audited FP8 base profile comparison, custom code hash comparison, and direct safetensors header metadata", "config_compatible": false, "notes": "The API metadata and model card identify MiniMaxAI/MiniMax-M2.7 as the base model. Manual comparison against the audited FP8 base profile and lukealonso NVFP4 package found identical custom MiniMaxM2 configuration/modeling code hashes and matching tensor, MoE, attention, rope, and full-context behavior. The served NVIDIA config records max_position_embeddings 196608 while the official FP8 base profile records 204800, so this profile uses the served NVIDIA config directly." }, "architecture": { "canonical_architecture_id": "minimax-m2-7", "max_context_tokens": 196608, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 139.862709248, "main_resident_weight_gb": 138.633516032, "auxiliary_resident_weight_gb": 1.229193216, "fixed_weight_gb": 12.250366976, "routed_expert_weight_gb": 0.493684176, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_modelopt_nvfp4_bf16_f8_f32_u8", "traffic_scope": "ordinary text decode through non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header data_offsets are used as the byte source of truth for the 15 indexed model shards. Expert tensors include ModelOpt NVFP4 U8 payloads plus F8_E4M3 and F32 scale tensors. Per-expert bytes are uniform at 493684176 bytes, so routed_expert_weight_gb is the exact routed expert byte sum divided by 256. Fixed traffic includes non-expert layer tensors, model.norm.weight, and the separate lm_head.weight." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null, matching the audited MiniMax M2 full-context behavior. This profile charges full-context K and V streams for all 62 language layers. NVIDIA's hf_quant_config records FP8 KV cache quantization, so KV storage/read bytes are one byte per scalar." }, "notes": "MiniMaxM2ForCausalLM MoE profile using the served NVIDIA custom config, ModelOpt quantization config, audited base/code comparison, and direct safetensors header grouping." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.6115456065076326, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-sglang-modelopt-nvfp4-moe-fp8-kv-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4/BF16/F8/F32/U8 safetensors bytes plus FP8 KV bytes. NVFP4 dequantization, activation quantization, router compute, expert compute, parallel communication, cache locality, and writes are outside this memory-side bound.", "notes": "The served config/hf_quant_config quantize Linear targets with 4-bit float weights and activations at group size 16 while ignoring lm_head, every MoE gate, and every self_attn module. hf_quant_config records kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA MiniMax M2.7 NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/MiniMax-M2.7-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "safetensors_dtype_split", "commit_sha" ], "notes": "At commit e79701cb1f9dce8fe5395b9ed2b20170beebecde, the API reports a public non-gated NVIDIA Software and Model Evaluation License safetensors repo with ModelOpt, minimax_m2, MiniMax, quantized, NVFP4/nvfp4, 8-bit, deploy:azure, and region:us tags. Current downloads were 100248 when audited. The API safetensors block reports BF16 1278796288, F32 2730491904, F8_E4M3 14042529792, U8 112340238336, and parameters.total 116349510656." }, { "label": "NVIDIA MiniMax M2.7 NVFP4 model card", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "serving" ], "notes": "The card describes this as an NVFP4-quantized MiniMaxAI/MiniMax-M2.7 artifact produced with nvidia-modelopt v0.43.0. It lists Sparse MoE architecture, 230B total parameters, 10B active parameters, 62 layers, hidden size 3072, 256 local experts, 8 experts per token, 204800 input context length, and SGLang/vLLM deployment guidance for Blackwell systems." }, { "label": "NVIDIA MiniMax M2.7 NVFP4 served config", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/raw/e79701cb1f9dce8fe5395b9ed2b20170beebecde/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records MiniMaxM2ForCausalLM, minimax_m2, dtype bfloat16, hidden size 3072, intermediate size 1536, 62 layers, 48 attention heads, 8 KV heads, head_dim 128, 256 local experts, 8 experts per token, no shared expert, max_position_embeddings 196608, sliding_window null, rope_theta 5000000, rotary_dim 64, sigmoid routing with routing bias, use_qk_norm true, tie_word_embeddings false, use_mtp true, and ModelOpt NVFP4 quantization metadata." }, { "label": "NVIDIA MiniMax M2.7 NVFP4 hf_quant_config", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/raw/e79701cb1f9dce8fe5395b9ed2b20170beebecde/hf_quant_config.json", "source_type": "config", "supports": [ "serving", "weight_format", "kv_store_format", "kv_read_format", "quantized_module_scope" ], "notes": "hf_quant_config records producer modelopt 0.43.0rc2.dev96+g3baa2da62.d20260410, quant_algo NVFP4, kv_cache_quant_algo FP8, 4-bit float weights and input activations with group_size 16, and exclude_modules for lm_head, every model.layers.*.block_sparse_moe.gate, and every model.layers.*.self_attn* module." }, { "label": "MiniMax M2.7 base/code comparison", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/d494266a4affc0d2995ba1fa35c8481cbd84294b/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found the NVIDIA artifact, MiniMaxAI FP8 base, and lukealonso NVFP4 package use identical configuration_minimax_m2.py and modeling_minimax_m2.py code hashes. Checked config fields match for model type, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, expert count, experts per token, rope theta, QK norm, and full-context sliding_window behavior. The NVIDIA and lukealonso NVFP4 repos set max_position_embeddings 196608 while the FP8 base config sets 204800." }, { "label": "NVIDIA MiniMax M2.7 NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/resolve/e79701cb1f9dce8fe5395b9ed2b20170beebecde/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The resolved safetensors index maps 191087 tensors across 15 shards. Range-read headers found 139.862709248 GB of tensor payload and 0.024738656 GB of safetensors header bytes. Payload dtypes are U8 112.340238336 GB, F8_E4M3 14.042529792 GB, F32 10.922348544 GB, and BF16 2.557592576 GB. model.embed_tokens.weight contributes 1.229193216 GB resident-only. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Non-expert layer tensors, model.norm.weight, and lm_head.weight total 12.250366976 GB. Routed expert tensors total 126.383149056 GB, averaging 0.493684176 GB per expert across 256 experts." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served config, hf_quant_config, audited MiniMax M2.7 base profile comparison, custom code hash comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile models ordinary text decode for NVIDIA's NVFP4 package. It uses the served 196608-token config rather than the base model card's 204800-token context and charges FP8 KV because NVIDIA's ModelOpt metadata explicitly declares FP8 KV-cache quantization." }, { "id": "nvidia--music-flamingo-2601-hf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/music-flamingo-2601-hf", "title": "NVIDIA Music Flamingo 2601 HF", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA Music Flamingo's BF16 Qwen2-style audio-language package.", "model_family": "music-flamingo-qwen2-audio", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B", "relation": "derived_package", "source": "Model card architecture notes, served MusicFlamingo config, Qwen2 text_config geometry, and direct safetensors header grouping", "config_compatible": false, "notes": "The model card says Music Flamingo uses a decoder-only Qwen2.5-7B backbone. The served package is not a drop-in Qwen2.5-7B checkpoint: it wraps the text model in MusicFlamingoForConditionalGeneration, changes vocabulary/context fields, adds audio tokens, an AF-Whisper audio encoder, and a multimodal projector. This profile therefore audits the served package directly and uses Qwen2.5 only as lineage context." }, "architecture": { "canonical_architecture_id": "music-flamingo-qwen2", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.26721536, "swept_params_b": 7.069214208, "auxiliary_resident_params_b": 1.198001152, "resident_weight_gb": 16.53443072, "swept_weight_gb": 14.138428416, "auxiliary_resident_weight_gb": 2.396002304, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes language_model.model.embed_tokens.weight input lookup and includes language_model layers, language_model.model.norm.weight, and language_model.lm_head.weight output projection", "auxiliary_scope": "language_model.model.embed_tokens.weight, audio_tower, and multi_modal_projector are resident for token/audio conditioning but are not swept as full matrices for each ordinary generated text token", "notes": "Range-read safetensors headers record only BF16 tensors. The text config marks tie_word_embeddings false and the checkpoint stores separate language_model.model.embed_tokens.weight and language_model.lm_head.weight tensors, so the input embedding is resident-only for ordinary decode while lm_head remains swept." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text_config records 28 full-attention Qwen2 layers, 4 KV heads, hidden_size 3584, 28 attention heads, sliding_window null, and no sliding-attention layers. The 128 head dimension is derived from hidden_size divided by attention heads. Music/audio encoder prefill and audio feature extraction are outside this ordinary text-decode profile." }, "notes": "MusicFlamingoForConditionalGeneration combines an AF-Whisper audio encoder, a multimodal projector, and a Qwen2-style decoder. This profile estimates memory-side ordinary text decode after any audio/text prefill, not audio encoder throughput, audio feature extraction, or projector prefill." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-musicflamingo-qwen2-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The served config records dtype bfloat16 and the safetensors header records only BF16 tensors. The model card shows Transformers generation examples and optional static-cache generation configuration; this profile uses BF16 full-context text KV for ordinary decode." }, "evidence": [ { "label": "NVIDIA Music Flamingo API metadata", "url": "https://huggingface.co/api/models/nvidia/music-flamingo-2601-hf", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "weight_format", "total_params_b" ], "notes": "At repo SHA 6b5be086d52f65a1e204cb0faf70bf54e2741ecd, the API records a public non-gated Transformers audio-text-to-text repo with musicflamingo, text2text-generation, music/songs, music understanding, music reasoning, dataset:nvidia/MF-Skills, arxiv:2511.10289, arxiv:2505.13032, license other, endpoints_compatible, and region:us tags. Current downloads are 188,844. The API safetensors block reports BF16 8,267,215,360 parameters." }, { "label": "NVIDIA Music Flamingo model card", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "audio_scope", "serving" ], "notes": "The model card says Music Flamingo is an audio-language model for music understanding, processes audio in 30-second windows with a 20-minute cap, supports text and audio inputs, uses Audio Flamingo 3, an AF-Whisper unified audio encoder, an MLP-based audio adaptor, and a decoder-only Qwen2.5-7B LLM backbone. The card lists max text input length 24,000 tokens and max text output length 2,048 tokens, and shows Transformers generation examples." }, { "label": "NVIDIA Music Flamingo config", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/raw/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "audio_scope", "embedding_layout" ], "notes": "The config records MusicFlamingoForConditionalGeneration / model_type musicflamingo, dtype bfloat16, a text_config with model_type qwen2, 28 full-attention layers, hidden_size 3584, intermediate_size 18944, 28 attention heads, 4 KV heads, max_position_embeddings 32768, model_max_length 32768, sliding_window null, use_sliding_window false through all-full layer_types, tie_word_embeddings false, and vocab_size 151672. The audio_config records audioflamingo3_encoder with 32 layers, hidden_size 1280, 20 attention heads, intermediate_size 5120, 128 mel bins, and max_source_positions 1500." }, { "label": "NVIDIA Music Flamingo processor and generation configs", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/raw/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/processor_config.json", "source_type": "config", "supports": [ "audio_scope", "serving" ], "notes": "processor_config records MusicFlamingoProcessor, WhisperFeatureExtractor, 16 kHz audio, 30-second chunks, max_audio_len 1200, audio token markers, and 128 mel bins. generation_config records max_new_tokens 2048 and the special text/audio token ids." }, { "label": "NVIDIA Music Flamingo safetensors header", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/resolve/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The linked object is 16.534531504 GB with a 100,776-byte safetensors header. Range-reading the header found 830 BF16 tensors totaling 8.267215360B parameters / 16.534430720 GB direct tensor payload. language_model tensors total 15.225613312 GB. language_model.model.embed_tokens.weight is BF16 [151672, 3584] and contributes 0.543592448B parameters / 1.087184896 GB resident-only for ordinary decode. language_model.lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Text layer tensors plus final norm plus lm_head total 7.069214208B parameters / 14.138428416 GB. audio_tower tensors total 1.273937920 GB, and multi_modal_projector tensors total 0.034879488 GB. Auxiliary resident tensors, defined as input embedding plus audio tower plus multimodal projector, total 1.198001152B parameters / 2.396002304 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served MusicFlamingo config, processor and generation configs, linked-object HEAD metadata, direct single-file safetensors header grouping, and Qwen2.5 7B lineage comparison." }, "notes": "This profile is for ordinary text-decode profile-backed bounds. It intentionally excludes audio encoder throughput, feature extraction, projector prefill, and end-to-end audio prompt latency." }, { "id": "nvidia--nemotron-3-nano-omni-30b-a3b-reasoning-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "title": "NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron 3 Nano Omni hybrid Mamba2-Transformer MoE multimodal serving package.", "model_family": "nemotron-h-nano-omni-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "relation": "derived_package", "source": "Model card network-architecture section plus direct language config comparison", "config_compatible": true, "notes": "The model card identifies NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 as the 31B A3B language backbone and adds CRADIO vision, Parakeet speech, projection, and multimodal wrapper modules. Manual config comparison found matching LLM geometry fields between this package's llm_config and the standalone BF16 backbone. The standalone text config explicitly sets mamba_ssm_cache_dtype float32, while this Omni BF16 wrapper omits that field; the local runtime code therefore allocates Mamba cache in the passed BF16 dtype for this package." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-omni-30b-a3b-reasoning", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 66.03127052, "main_resident_weight_gb": 62.451243392, "auxiliary_resident_weight_gb": 3.580027128, "fixed_weight_gb": 3.701627264, "routed_expert_weight_gb": 0.458981376, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32_i64", "traffic_scope": "ordinary text decode through language_model excluding resident-only input embedding; fixed traffic includes Mamba, attention, shared experts, routers, norms, and lm_head tensors", "auxiliary_scope": "language_model.backbone.embeddings.weight, vision_model tensors, sound_encoder tensors, sound_projection tensors, and mlp1 tensors are resident for the package but not swept as full matrices for each generated text token", "shared_expert_notes": "The config records one shared expert and six routed experts per token. The v1 adapter folds always-on shared expert traffic into the fixed active-weight term.", "notes": "Header-derived stored bytes are used instead of rounded 31B/3B card parameters. Routed expert tensors are byte-uniform across all 128 expert indexes and 23 MoE layers. The package is almost entirely BF16, with tiny F32 router correction tensors and I64 audio batch-norm counters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid pattern maps '*' to attention layers at indexes 5, 12, 19, 26, 33, and 42. Attention stores K and V before repeating two KV heads across 32 query heads." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.024870912, "read_gb_per_output_token": 0.024870912, "state_formula": "23 Mamba layers * ((4096 intermediate scalars * 4 convolution kernel) + (4096 intermediate scalars * 128 SSM state)) * 2 BF16 bytes", "notes": "The custom NemotronHHybridDynamicCache initializes conv_states and ssm_states with the dtype passed to the cache. This BF16 Omni config omits mamba_ssm_cache_dtype, and the model card says to omit FP8 KV cache for BF16, so this profile charges BF16 Mamba state. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for six attention layers plus fixed BF16 Mamba recurrent state for 23 Mamba layers." }, "notes": "This profile models ordinary text decode after any multimodal prefill. Vision, audio, video preprocessing, and multimodal prefill throughput are outside Bounds Engine v1. The card advertises 256k context, the top wrapper default is 131072, and the LLM config records 262144 max position embeddings; this profile uses the LLM text-decode context ceiling." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "bf16-attention-bf16-mamba-state", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-attention-bf16-mamba-state", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-bf16-hybrid-mamba-moe-omni", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes, BF16 attention KV, and BF16 Mamba recurrent state traffic. Compute, router cost, activation traffic, state writes, multimodal encoder compute, and prefill traffic are outside Bounds Engine v1.", "notes": "The card's vLLM command explicitly says to omit --kv-cache-dtype fp8 for BF16. The 17 safetensors shards store BF16 tensors plus tiny F32 and I64 tensors; weight_bytes_per_param remains the nominal BF16 byte width, while exact resident and traffic fields are authoritative." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano Omni BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "At repo SHA 24e67ea000b7c2837fc8f9488aa2008524fac8ba, the HF API records a public transformers any-to-any repo with custom code, NVIDIA Open Model Agreement metadata, 1003984 downloads, region:us, and safetensors parameters BF16 33015629264 and F32 2950. The card states 31B total parameters, about 3B active parameters per token, 256k max context, video/audio/image/text inputs, text output, BF16 minimum hardware of one H100 80GB, and a vLLM command whose KV-cache comment says to omit FP8 KV for BF16." }, { "label": "NVIDIA Nemotron 3 Nano Omni BF16 config", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/24e67ea000b7c2837fc8f9488aa2008524fac8ba/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens" ], "notes": "The config records NemotronH_Nano_Omni_Reasoning_V3 with llm_config model_type nemotron_h, 52 layers, hybrid_override_pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME, 23 Mamba layers, 23 MoE layers, 6 attention layers, hidden_size 2688, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 128 routed experts, 6 routed experts per token, 1 shared expert, mamba_num_heads 64, mamba_head_dim 64, conv_kernel 4, ssm_state_size 128, BF16 dtype, and no mamba_ssm_cache_dtype override." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 backbone config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited language geometry between the Omni llm_config and the standalone BF16 backbone: 52-layer hybrid pattern, hidden size, intermediate sizes, attention geometry, MoE expert counts, Mamba geometry, 262144 max positions, vocabulary size, and untied embeddings. The standalone backbone additionally records mamba_ssm_cache_dtype float32, while the Omni BF16 wrapper omits that cache override." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/24e67ea000b7c2837fc8f9488aa2008524fac8ba/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. NemotronHHybridDynamicCache allocates attention key/value lists plus Mamba conv_states with shape batch_size x 6144 x 4 and ssm_states with shape batch_size x 64 x 64 x 128 for each Mamba layer, using the dtype passed to the cache. Attention projections store K and V with num_key_value_heads 2 and head_dim 128 before repeating to query heads." }, { "label": "NVIDIA Nemotron 3 Nano Omni BF16 safetensors headers", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/resolve/24e67ea000b7c2837fc8f9488aa2008524fac8ba/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "All 17 safetensors shard headers were range-read directly. The headers match the index tensor count of 7349 and index total_size 66.031270520 GB. Stored tensor bytes split into BF16 66.031258528 GB, F32 0.000011800 GB, and I64 0.000000192 GB. language_model tensors total 63.155886464 GB; language_model.backbone.embeddings.weight is 0.704643072 GB and resident-only for ordinary decode. Auxiliary resident tensors, defined as input embedding plus vision_model, sound_encoder, sound_projection, and mlp1 tensors, sum to 3.580027128 GB. Ordinary text resident tensors excluding input embedding sum to 62.451243392 GB. Fixed ordinary text traffic excluding routed experts and input embedding sums to 3.701627264 GB. Routed expert tensors sum to 58.749616128 GB, exactly 0.458981376 GB per expert index across all 23 MoE layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned BF16 Omni config, standalone BF16 backbone config comparison, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, multimodal resident bytes, BF16 KV/state assumptions, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nemotron-3-nano-omni-30b-a3b-reasoning-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4", "title": "NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's NVFP4 Nemotron 3 Nano Omni hybrid Mamba2-Transformer MoE serving artifact.", "model_family": "nemotron-h-nano-omni-hybrid-moe", "base_model_proof": { "base_model": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "relation": "quantized", "source": "Hugging Face base_model metadata plus direct served config comparison", "config_compatible": true, "notes": "The NVFP4 repo declares the BF16 repo as its quantized base. Manual comparison found matching top-level wrapper, LLM, audio, and vision architecture fields relevant to text-decode bounds; the NVFP4 repo adds ModelOpt mixed-precision quantization metadata." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-omni-30b-a3b-reasoning", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 22.40548488, "main_resident_weight_gb": 18.825524368, "auxiliary_resident_weight_gb": 3.579960512, "fixed_weight_gb": 2.302147728, "routed_expert_weight_gb": 0.12908888, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through language_model excluding resident-only input embedding; fixed traffic includes Mamba, attention, shared experts, routers, norms, and lm_head tensors", "auxiliary_scope": "language_model.backbone.embeddings.weight, vision_model tensors, sound_encoder tensors, sound_projection tensors, and mlp1 tensors are resident for the package but not swept as full matrices for each generated text token", "shared_expert_notes": "The config records one shared expert and six routed experts per token. The v1 adapter folds always-on shared expert traffic into the fixed active-weight term.", "notes": "Header-derived stored bytes are used instead of rounded 31B/3B model-card parameters. Routed expert tensors are byte-uniform across all 128 expert indexes and 23 MoE layers. The package mixes packed U8 NVFP4 expert tensors, FP8 non-routed language tensors, BF16 embeddings/multimodal tensors, and F32 scale/state tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid pattern maps '*' to attention layers at indexes 5, 12, 19, 26, 33, and 42. Attention stores K and V before repeating two KV heads across 32 query heads." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.024870912, "read_gb_per_output_token": 0.024870912, "state_formula": "23 Mamba layers * ((4096 intermediate scalars * 4 convolution kernel) + (4096 intermediate scalars * 128 SSM state)) * 2 BF16 bytes", "notes": "The custom HybridMambaAttentionDynamicCache initializes conv_states and ssm_states for Mamba layers with the model dtype. This profile charges one full fixed Mamba state read per generated token; compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context K/V for six attention layers plus fixed BF16 Mamba recurrent state for 23 Mamba layers." }, "notes": "This profile models ordinary text decode after any multimodal prefill. Vision, audio, video preprocessing, and multimodal prefill throughput are outside Bounds Engine v1. The model card advertises 256k context and the LLM config records 262144 max position embeddings; the card's vLLM examples start at --max-model-len 131072 as a deployment/memory-tuning setting." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-hybrid-moe", "dequantization_notes": "The memory-side bound charges stored safetensors bytes and FP8 KV bytes. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, and Mamba state writes are outside Bounds Engine v1.", "notes": "The ModelOpt recipe is mixed precision: routed expert up/down projections use NVFP4, non-routed language projections and shared experts use FP8, and vision/audio encoders plus projection modules stay BF16. weight_bytes_per_param records the nominal NVFP4 payload; the audited adapter uses exact stored tensor bytes for resident and per-token weight traffic." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano Omni NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b", "max_context_tokens", "serving" ], "notes": "The HF API records repo SHA dc5f0b0bfddf8b6e0f5891475be9af05b80126fe, any-to-any pipeline, custom transformers code, base_model nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 with base_model:quantized metadata, ModelOpt/NVFP4 tags, and safetensors dtype groups. The card states 31B total parameters, about 3B active parameters per token, 256k max context, minimum NVFP4 hardware of RTX 5090 32GB with DGX Spark and Jetson Thor also supported, and vLLM serving with --kv-cache-dtype fp8." }, { "label": "NVIDIA Nemotron 3 Nano Omni NVFP4 config", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/raw/dc5f0b0bfddf8b6e0f5891475be9af05b80126fe/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens" ], "notes": "The config records NemotronH_Nano_Omni_Reasoning_V3 with llm_config model_type nemotron_h, 52 layers, hybrid_override_pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME, 23 Mamba layers, 23 MoE layers, 6 attention layers, hidden_size 2688, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 128 routed experts, 6 routed experts per token, 1 shared expert, mamba_num_heads 64, mamba_head_dim 64, conv_kernel 4, ssm_state_size 128, and quantization_config kv_cache_scheme float 8-bit." }, { "label": "NVIDIA Nemotron 3 Nano Omni BF16 base config comparison", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/24e67ea000b7c2837fc8f9488aa2008524fac8ba/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top model type, top max_sequence_length 131072, LLM model type, 52-layer hybrid pattern, hidden size, intermediate sizes, expert counts, routed experts per token, attention head geometry, 262144 LLM max_position_embeddings, Mamba state geometry, and audio/vision geometry between the BF16 base and NVFP4 repo." }, { "label": "NVIDIA Nemotron 3 Nano Omni NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/raw/dc5f0b0bfddf8b6e0f5891475be9af05b80126fe/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 0.43.0rc2.dev117, kv_cache_quant_algo FP8, quant_algo MIXED_PRECISION, 98 FP8 quantized non-routed language layers, and 5888 NVFP4 quantized routed expert projection entries." }, { "label": "NVIDIA Nemotron 3 Nano Omni NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/resolve/dc5f0b0bfddf8b6e0f5891475be9af05b80126fe/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The three safetensors headers were range-read directly. Stored tensor bytes sum to 22.40548488 GB across 25205 tensors: BF16 4.435134336 GB, F32 0.031713552 GB, F8_E4M3 3.251232768 GB, I64 0.000000192 GB, and U8 14.687404032 GB. Ordinary text resident tensors excluding input embedding sum to 18.825524368 GB. Auxiliary resident tensors, defined as language_model.backbone.embeddings.weight plus vision_model, sound_encoder, sound_projection, and mlp1 tensors, sum to 3.579960512 GB. Fixed ordinary text traffic excluding routed experts and input embedding sums to 2.302147728 GB. Routed expert tensors sum to 16.52337664 GB, exactly 0.12908888 GB per expert index across all 23 MoE layers." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/raw/dc5f0b0bfddf8b6e0f5891475be9af05b80126fe/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. HybridMambaAttentionDynamicCache allocates attention key/value lists plus Mamba conv_states with shape batch_size x 4096 x 4 and ssm_states with shape batch_size x 4096 x 128 for each Mamba layer. Attention projections store K and V with num_key_value_heads 2 and head_dim 128 before repeating to query heads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the served NVFP4 config, BF16 base config comparison, ModelOpt quantization config, model card serving notes, custom runtime code, and direct safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed NVFP4/FP8 weight traffic, or FP8 KV assumptions from the repo name." }, { "id": "nvidia--nemotron-labs-diffusion-3b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Nemotron-Labs-Diffusion-3B-Base", "title": "NVIDIA Nemotron Labs Diffusion 3B Base", "summary": "Audited memory-side ordinary AR text-decode bounds profile for the BF16 Nemotron Labs Diffusion 3B Base custom-code checkpoint.", "model_family": "nemotron-labs-diffusion-dense-ar", "architecture": { "canonical_architecture_id": "nemotron-labs-diffusion-3b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.83165952, "swept_params_b": 3.429006336, "auxiliary_resident_params_b": 0.402653184, "resident_weight_gb": 7.66331904, "swept_weight_gb": 6.858012672, "auxiliary_resident_weight_gb": 0.805306368, "resident_parameter_scope": "Direct safetensors tensor-header element count for the complete BF16 checkpoint", "swept_parameter_scope": "encoder layers, encoder norm, and untied diffusion_head output projection used by ar_generate", "auxiliary_scope": "encoder.embed_tokens.weight is resident for token lookup but not swept as a full matrix per generated text token", "notes": "This profile models the custom ar_generate path only. The separate untied diffusion_head.weight is the output projection and remains swept during ordinary AR decode, while encoder.embed_tokens.weight is resident-only for decode." }, "kv_adapter": { "kind": "full_context", "layers": 26, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records no sliding_window, 26 layers, 8 KV heads, and 128 head dimension. Runtime review found ar_generate creates a DynamicCache and calls the Ministral encoder with use_cache=true and causal masking; the attention module updates cached key and value states before attention." }, "notes": "Nemotron-Labs-Diffusion is a tri-mode language model. Bounds Engine v1 only profiles ordinary autoregressive text decode through ar_generate; it does not model diffusion parallel decoding, block diffusion sampling, or linear self-speculation speedups." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "custom-nemotron-labs-diffusion-ar-bf16-text-decode-memory-bound", "notes": "The checkpoint stores all tensors as BF16. The memory-side bound charges the ordinary AR path implemented by ar_generate, not the model card's diffusion or self-speculation modes." }, "evidence": [ { "label": "NVIDIA Nemotron Labs Diffusion 3B Base model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-3B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "weight_format", "total_params_b" ], "notes": "At commit 4bd78816ba988ca8b71b8f432b00b682572b236c, the API records a public non-gated NVIDIA Open Model License repo with transformers, safetensors, nemotron_labs_diffusion, custom_code, text-generation, conversational, and region:us tags; current downloads were 270224. The model card describes a tri-mode language model supporting AR decoding, diffusion-based parallel decoding, and self-speculation by switching attention patterns." }, { "label": "NVIDIA Nemotron Labs Diffusion 3B Base config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The pinned config records NemotronLabsDiffusionModel, custom AutoModel code, torch_dtype bfloat16, dlm_paradigm bidirectional, use_cache false by default, 26 layers, hidden size 3072, intermediate size 9216, 32 attention heads, 8 KV heads, head_dim 128, no sliding_window, max_position_embeddings 4096, vocab_size 131072, and tie_word_embeddings false." }, { "label": "Nemotron Labs Diffusion custom AR runtime review", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/modeling_nemotron_labs_diffusion.py", "source_type": "manual_review", "supports": [ "decode_policy_scope", "kv_cache", "weight_traffic_scope" ], "notes": "Manual review found ar_generate explicitly sets each attention layer's diffusion_lm flag to false, creates a DynamicCache, performs the prompt pass with use_cache=true, and then decodes one token at a time through self.encoder plus diffusion_head. The method docstring says it bypasses diffusion-specific code paths so KV cache and causal masking behave like MistralForCausalLM/vLLM." }, { "label": "Nemotron Labs Diffusion Ministral attention runtime review", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/modeling_ministral.py", "source_type": "manual_review", "supports": [ "kv_cache", "layers", "full_context" ], "notes": "Manual review found Ministral3Attention builds query, key, and value projections with num_key_value_heads x head_dim K/V tensors, updates past_key_values when use_cache is true, and uses create_causal_mask when ar_generate calls the encoder with causal masking. Because the served config has sliding_window null, the AR profile charges full-context KV." }, { "label": "NVIDIA Nemotron Labs Diffusion 3B Base safetensors header", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "Range-reading the safetensors header found a 27,952-byte header and 237 BF16 tensors totaling 7.663319040 GB across 3,831,659,520 elements. Tensor buckets are encoder.embed_tokens.weight 0.805306368 GB, diffusion_head.weight 0.805306368 GB, encoder.layers.* 6.052700160 GB, and encoder.norm.weight 0.000006144 GB. The ordinary AR swept path is encoder layers plus norm plus diffusion_head, totaling 6.858012672 GB, while encoder.embed_tokens.weight is resident-only." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned config, custom runtime code review, and direct safetensors header byte grouping." }, "notes": "This profile is intentionally limited to ordinary AR text decode through the custom ar_generate method. Do not use it as evidence for diffusion parallel decoding, block diffusion, or linear self-speculation throughput." }, { "id": "nvidia--nemotron-labs-diffusion-8b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/Nemotron-Labs-Diffusion-8B-Base", "title": "NVIDIA Nemotron-Labs-Diffusion 8B Base BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 Nemotron-Labs-Diffusion 8B Base tri-mode repo.", "model_family": "nemotron-labs-diffusion-dense", "architecture": { "canonical_architecture_id": "nemotron-labs-diffusion-8b-base", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.48955392, "swept_params_b": 7.952683008, "auxiliary_resident_params_b": 0.536870912, "resident_weight_gb": 16.97910784, "swept_weight_gb": 15.905366016, "auxiliary_resident_weight_gb": 1.073741824, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "encoder.layers plus encoder.norm plus diffusion_head.weight; recorded for AR-mode evidence only because production bounds are disabled", "auxiliary_scope": "encoder.embed_tokens.weight is resident but is not swept as a full matrix for ordinary AR token decode", "notes": "The single safetensors header is exact: 309 BF16 tensors total 16.979107840 GB. encoder.layers tensors total 14.831616000 GB, encoder.norm is 0.000008192 GB, diffusion_head.weight is 1.073741824 GB, and encoder.embed_tokens.weight is 1.073741824 GB. Bounds Engine v1 does not turn the AR-mode swept bytes into production throughput for this repo because the served model exposes multiple generation modes without a workload mode selector." }, "kv_adapter": { "kind": "unknown", "reason": "Nemotron-Labs-Diffusion is a tri-mode language model: the custom repo exposes ordinary AR decoding, block-wise diffusion decoding, and linear self-speculation. The default model generate path performs block-wise diffusion with repeated masked-token denoising and optional causal context, while ar_generate is a separate explicit mode. Bounds Engine v1 only has one profile per repo and no generation-mode selector, so publishing a normal AR tok/s bound would be misleading for this multi-mode repo.", "notes": "The AR helper path would use BF16 full-context KV with 34 layers, 8 KV heads, and 128 head dimension, equal to 0.139264 GB per 1K tokens. A production profile needs a mode-dispatched adapter that separately models AR, diffusion block length and denoising evaluations, and self-speculation draft/verify acceptance behavior." }, "notes": "This profile intentionally fails closed even though config, AR geometry, and tensor headers are accessible, because the repo's primary value and default generation API are not represented by ordinary one-token autoregressive decode alone." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-nemotron-labs-diffusion-tri-mode", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The API safetensors block and direct single-file safetensors header read record only BF16 tensors. Production KV/state bounds are disabled until the comparison engine can select and model the generation mode." }, "evidence": [ { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-8B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 59ff0ffee284112fc6ccf37493ced41a11031434, the live API records a public non-gated text-generation Transformers repo with custom_code, nemotron_labs_diffusion, region:us, current downloads 479470, NVIDIA Nemotron Open Model License metadata, and BF16 safetensors parameters 8489553920." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/raw/59ff0ffee284112fc6ccf37493ced41a11031434/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "layers", "kv_heads", "head_dim", "serving", "unsupported_reason" ], "notes": "The config records NemotronLabsDiffusionModel, model_type nemotron_labs_diffusion, bfloat16 dtype, dlm_paradigm bidirectional, block_size 32, use_cache false, 34 hidden layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, 4096 max position embeddings, untied embeddings, mask_token_id 100, and YaRN RoPE parameters." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base generation config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/raw/59ff0ffee284112fc6ccf37493ced41a11031434/generation_config.json", "source_type": "config", "supports": [ "serving", "unsupported_reason" ], "notes": "The generation config is inherited from the model config and records use_cache false, reinforcing that the repo's generic generation setup is not a normal cached CausalLM decode profile." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base model card", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The card describes the model as tri-mode, supporting AR decoding, diffusion-based parallel decoding, and linear self-speculation by switching attention patterns. Its example calls ar_generate for AR mode, generate with block_length and threshold for diffusion mode, and linear_spec_generate for self-speculation." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base custom modeling code", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/raw/59ff0ffee284112fc6ccf37493ced41a11031434/modeling_nemotron_labs_diffusion.py", "source_type": "manual_review", "supports": [ "unsupported_reason", "generation_algorithm", "kv_adapter" ], "notes": "Manual review found generate performs block-wise diffusion over mask-token blocks with repeated denoising forward evaluations, causal KV updates between blocks, and diffusion_lm attention-mode toggles. ar_generate separately forces diffusion_lm false and performs ordinary cached AR decoding. linear_spec_generate alternates diffusion draft passes with causal AR verification and cache cropping." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B Base safetensors header", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/resolve/59ff0ffee284112fc6ccf37493ced41a11031434/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format" ], "notes": "A direct range read of the single safetensors header found a 36872-byte header and 309 BF16 tensors totaling 8489553920 parameters / 16.979107840 GB. Tensor groups are diffusion_head.weight 1.073741824 GB, encoder.embed_tokens.weight 1.073741824 GB, encoder.layers 14.831616000 GB, and encoder.norm 0.000008192 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from live HF API metadata, model card, pinned config, generation config, custom modeling code, and direct single-file safetensors header range read. Marked unsupported because Bounds Engine v1 lacks a mode-dispatched adapter for tri-mode AR/diffusion/self-speculation repos." }, "unsupported_reason": "Nemotron-Labs-Diffusion-8B-Base exposes AR, diffusion, and self-speculation generation modes from the same checkpoint, and its default custom generate path is block-wise diffusion rather than ordinary one-token autoregressive decode. Bounds Engine v1 cannot select or model those modes separately, so production tok/s bounds are disabled to avoid misleading comparisons.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine has explicit workload-mode profiles for AR, diffusion block decoding, and linear self-speculation." }, { "id": "nvidia--nemotron-labs-diffusion-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "nvidia/Nemotron-Labs-Diffusion-8B", "title": "NVIDIA Nemotron-Labs-Diffusion 8B BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the BF16 Nemotron-Labs-Diffusion 8B tri-mode repo.", "model_family": "nemotron-labs-diffusion-dense", "architecture": { "canonical_architecture_id": "nemotron-labs-diffusion-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.48955392, "swept_params_b": 7.952683008, "auxiliary_resident_params_b": 0.536870912, "resident_weight_gb": 16.97910784, "swept_weight_gb": 15.905366016, "auxiliary_resident_weight_gb": 1.073741824, "resident_parameter_scope": "safetensors_header_stored_bf16_core_checkpoint", "swept_parameter_scope": "encoder.layers plus encoder.norm plus diffusion_head.weight; recorded for AR-mode evidence only because production bounds are disabled", "auxiliary_scope": "encoder.embed_tokens.weight is resident but is not swept as a full matrix for ordinary AR token decode", "notes": "The single core safetensors header is exact: 309 BF16 tensors total 16.979107840 GB. encoder.layers tensors total 14.831616000 GB, encoder.norm is 0.000008192 GB, diffusion_head.weight is 1.073741824 GB, and encoder.embed_tokens.weight is 1.073741824 GB. The optional linear_spec_lora/adapter_model.safetensors file stores 0.142606336 GB of F32 LoRA tensors and is not included in the core checkpoint resident/swept fields. Bounds Engine v1 does not turn the AR-mode swept bytes into production throughput for this repo because the served model exposes multiple generation modes without a workload mode selector." }, "kv_adapter": { "kind": "unknown", "reason": "Nemotron-Labs-Diffusion is a tri-mode language model: the custom repo exposes ordinary AR decoding, block-wise diffusion decoding, and linear self-speculation. The default model generate path performs block-wise diffusion with repeated masked-token denoising and optional causal context, while ar_generate is a separate explicit mode and linear_spec_generate adds diffusion draft plus AR verification. Bounds Engine v1 only has one profile per repo and no generation-mode selector, so publishing a normal AR tok/s bound would be misleading for this multi-mode repo.", "notes": "The AR helper path would use BF16 full-context KV with 34 layers, 8 KV heads, and 128 head dimension, equal to 0.139264 GB per 1K tokens. A production profile needs a mode-dispatched adapter that separately models AR, diffusion block length and denoising evaluations, and self-speculation draft/verify acceptance behavior." }, "notes": "This profile intentionally fails closed even though config, AR geometry, custom code, and tensor headers are accessible, because the repo's primary value and default generation API are not represented by ordinary one-token autoregressive decode alone." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-nemotron-labs-diffusion-tri-mode", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The API safetensors block and direct single-file safetensors header read record only BF16 core checkpoint tensors. Production KV/state bounds are disabled until the comparison engine can select and model the generation mode." }, "evidence": [ { "label": "NVIDIA Nemotron-Labs-Diffusion 8B API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-8B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 16c67f0560b912e93e0cabb6e0c4f5c3086d95fc, the live API records a public non-gated text-generation Transformers repo with custom_code, nemotron_labs_diffusion, region:us, current downloads 125076, NVIDIA Nemotron Open Model License metadata from the card, and BF16 safetensors parameters 8489553920." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/raw/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/config.json", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "layers", "kv_heads", "head_dim", "serving", "unsupported_reason" ], "notes": "The config records NemotronLabsDiffusionModel, model_type nemotron_labs_diffusion, bfloat16 dtype, dlm_paradigm bidirectional, block_size 32, use_cache false, 34 hidden layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, untied embeddings, mask_token_id 100, and YaRN RoPE parameters with factor 16." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B generation config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/raw/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/generation_config.json", "source_type": "config", "supports": [ "serving", "unsupported_reason" ], "notes": "The generation config is inherited from the model config and records use_cache false, reinforcing that the repo's generic generation setup is not a normal cached CausalLM decode profile." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B model card", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B", "source_type": "model_card", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The card describes the model as tri-mode, supporting AR decoding, diffusion-based parallel decoding, and linear self-speculation by switching attention patterns. Its examples call ar_generate for AR mode, generate with block_length and threshold for diffusion mode, and linear_spec_generate for self-speculation. The card also documents an optional LoRA adapter for the linear self-speculation drafter." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B custom modeling code", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/raw/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/modeling_nemotron_labs_diffusion.py", "source_type": "manual_review", "supports": [ "unsupported_reason", "generation_algorithm", "kv_adapter" ], "notes": "Manual review found generate performs block-wise diffusion over mask-token blocks with repeated denoising forward evaluations, causal KV updates between blocks, and diffusion_lm attention-mode toggles. ar_generate separately forces diffusion_lm false and performs ordinary cached AR decoding. linear_spec_generate alternates diffusion draft passes with causal AR verification, optional LoRA toggling, and cache cropping." }, { "label": "NVIDIA Nemotron-Labs-Diffusion 8B safetensors headers", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/resolve/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "optional_lora_adapter_size" ], "notes": "A direct range read of the core model.safetensors header found a 36872-byte header and 309 BF16 tensors totaling 8489553920 parameters / 16.979107840 GB. Tensor groups are diffusion_head.weight 1.073741824 GB, encoder.embed_tokens.weight 1.073741824 GB, encoder.layers 14.831616000 GB, and encoder.norm 0.000008192 GB. A separate range read found the optional linear_spec_lora/adapter_model.safetensors file contains 68 F32 tensors totaling 0.142606336 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Reviewed from live HF API metadata, model card, pinned config, generation config, custom modeling code, and direct safetensors header range reads for the core checkpoint and optional self-spec LoRA adapter. Marked unsupported because Bounds Engine v1 lacks a mode-dispatched adapter for tri-mode AR/diffusion/self-speculation repos." }, "unsupported_reason": "Nemotron-Labs-Diffusion-8B exposes AR, diffusion, and self-speculation generation modes from the same checkpoint, and its default custom generate path is block-wise diffusion rather than ordinary one-token autoregressive decode. Bounds Engine v1 cannot select or model those modes separately, so production tok/s bounds are disabled to avoid misleading comparisons.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine has explicit workload-mode profiles for AR, diffusion block decoding, and linear self-speculation." }, { "id": "nvidia--nemotron-mini-4b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Nemotron-Mini-4B-Instruct", "title": "NVIDIA Nemotron Mini 4B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron Mini 4B Instruct repo.", "model_family": "nemotron-minitron-dense", "base_model_proof": { "base_model": "nvidia/Minitron-4B-Base", "relation": "finetune", "source": "Hugging Face model metadata, model card, and direct served config comparison", "config_compatible": true, "notes": "The model card identifies this repo as a fine-tuned version of nvidia/Minitron-4B-Base. Manual config comparison against the current base config found the same architecture, hidden size, layer count, attention geometry, context length, MLP size, RoPE settings, dtype, and vocabulary size; the only geometry spelling difference is base head_dim 128 versus instruct kv_channels 128." }, "architecture": { "canonical_architecture_id": "nemotron-mini-4b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.190509056, "swept_params_b": 3.404077056, "auxiliary_resident_params_b": 0.786432, "resident_weight_gb": 8.381018112, "swept_weight_gb": 6.808154112, "auxiliary_resident_weight_gb": 1.572864, "resident_parameter_scope": "pytorch_zip_storage_bf16_tensor_payload", "swept_parameter_scope": "ordinary text decode through model.layers.*, model.norm.*, and lm_head.weight, excluding resident-only input embeddings", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "notes": "Range-read PyTorch zip metadata and the checkpoint pickle record 324 BF16 tensor storages totaling 4190509056 parameters / 8.381018112 GB. The config records tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors, each [256000, 3072] / 1.572864 GB. Ordinary text decode sweeps the decoder layers, final norm/bias, and lm_head.weight; it excludes only the input embedding lookup. The .bin archive itself is 8.381129342 GB, with 0.000111230 GB of zip/pickle overhead that is not charged as accelerator-resident tensor memory." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 decoder layers, 8 KV heads, kv_channels 128, use_cache true, and 4096 max position embeddings. The profile charges ordinary full-context K and V cache as BF16 scalars." }, "notes": "NemotronForCausalLM is a dense decoder-only text model with grouped-query attention and RoPE. The model card says it supports 4096 tokens of context; the served config records max_position_embeddings 4096." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-nemotron-dense-text-decode", "dequantization_notes": "No quantized weight representation is assumed for this BF16 PyTorch checkpoint.", "notes": "The served config records torch_dtype bfloat16, and the PyTorch checkpoint pickle records only BFloat16Storage tensors. The model card's sample loads through Transformers without requesting KV quantization, so this profile keeps attention KV as BF16." }, "evidence": [ { "label": "Nemotron Mini 4B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Mini-4B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "base_model_proof", "model_family", "max_context_tokens" ], "notes": "At repo SHA 791833e92ebddb0bc2c1007f6d2b6764f886a2ae, the HF API records a public non-gated Transformers text-generation repo with PyTorch and NeMo artifacts, license:other, endpoints_compatible, region:us, and current downloads 390524. The model card describes the repo as a fine-tuned version of nvidia/Minitron-4B-Base, optimized for English roleplay/RAG/function-calling workloads, and states 4096-token context support." }, { "label": "Nemotron Mini 4B Instruct served config", "url": "https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct/raw/791833e92ebddb0bc2c1007f6d2b6764f886a2ae/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records NemotronForCausalLM, model_type nemotron, bfloat16, hidden_size 3072, intermediate_size 9216, 32 layers, 24 attention heads, 8 KV heads, kv_channels 128, max_position_embeddings 4096, partial_rotary_factor 0.5, tie_word_embeddings false, use_cache true, and vocab_size 256000." }, { "label": "Minitron 4B Base config comparison", "url": "https://huggingface.co/nvidia/Minitron-4B-Base/raw/3478fb8d5e0a4347e4ab08727f9f95f33f97df3b/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible" ], "notes": "The current base config has the same architecture and checked geometry as the instruct config. It uses head_dim 128 where the instruct config uses kv_channels 128; both describe the same KV head dimension." }, { "label": "Nemotron Mini 4B Instruct PyTorch checkpoint metadata", "url": "https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct/resolve/791833e92ebddb0bc2c1007f6d2b6764f886a2ae/pytorch_model.bin", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout", "lm_head_layout" ], "notes": "The linked PyTorch archive is 8.381129342 GB. A direct zip central-directory range read found 324 tensor storage entries totaling 8.381018112 GB, and a direct data.pkl range read found all storages are BF16. Tensor grouping gives model.layers.* 3.422011392 GB, model.norm.* 0.000012288 GB, lm_head.weight 1.572864000 GB, and model.embed_tokens.weight 1.572864000 GB." }, { "label": "Nemotron Mini 4B Instruct NeMo artifact size check", "url": "https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct/resolve/791833e92ebddb0bc2c1007f6d2b6764f886a2ae/nemo/nemotron-mini-4b-instruct.nemo", "source_type": "manual_review", "supports": [ "artifact_selection" ], "notes": "The alternate NeMo artifact is 8.386488320 GB. This profile targets the Transformers PyTorch checkpoint because the repo library is transformers and the model card's executable examples load AutoModelForCausalLM from this repo." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, Minitron base config comparison, linked artifact HEAD checks, direct PyTorch zip central-directory range read, and direct checkpoint pickle metadata extraction with fake storages." }, "notes": "This self-contained profile replaces the scraped 4B/8GB metadata estimate with exact BF16 tensor payload bytes and separates resident-only input embedding memory from per-token swept decode traffic." }, { "id": "nvidia--nvidia-nemotron-3-nano-30b-a3b-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "title": "NVIDIA Nemotron 3 Nano 30B A3B BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron 3 Nano 30B A3B hybrid Mamba2-Transformer MoE repo.", "model_family": "nemotron-h-nano-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16", "relation": "finetune", "source": "Model card training lineage plus direct served config comparison", "config_compatible": true, "notes": "The model card identifies the Base BF16 repo as the pretraining base. Manual comparison found matching audited geometry fields between the base and instruct configs; the instruct repo adds torch_dtype bfloat16 and mamba_ssm_cache_dtype float32 fields that do not change the bounds geometry." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 63.155892352, "main_resident_weight_gb": 62.45124928, "auxiliary_resident_weight_gb": 0.704643072, "fixed_weight_gb": 3.701633152, "routed_expert_weight_gb": 0.458981376, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding; fixed traffic includes Mamba, attention, router, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 13 shards and 6243 tensors totaling 63.155892352 GB: BF16 63.155868800 GB plus F32 0.000023552 GB. Routed expert tensors are byte-uniform across all 128 expert indexes and 23 MoE layers at exactly 0.458981376 GB per expert index. Non-routed ordinary decode traffic totals 3.701633152 GB, while the resident-only input embedding totals 0.704643072 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 5, 12, 19, 26, 33, and 42. Attention stores K and V with 2 KV heads and 128 head dimension before repeating to query heads." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.049364992, "read_gb_per_output_token": 0.049364992, "state_formula": "23 Mamba layers * (((4096 intermediate scalars + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel) * 2 BF16 bytes + (4096 intermediate scalars * 128 SSM state) * 4 F32 bytes)", "notes": "The config records mamba_ssm_cache_dtype float32 and the mixer convolution runs over the 6144-wide hidden+B+C projection. This profile charges BF16 convolution state plus F32 SSM recurrent state, one full fixed Mamba state read per generated token. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for six attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for twenty-three Mamba layers." }, "notes": "NemotronHForCausalLM uses a 52-layer hybrid pattern with 23 Mamba layers, 23 MoE layers, and 6 full-attention layers. The model card advertises support up to 1M context with VLLM_ALLOW_LONG_MAX_MODEL_LEN, while the default Hugging Face config and vLLM example use 262144; this profile uses the served config value." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "bf16-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-or-vllm-bf16-hybrid-mamba-moe", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes, BF16 attention KV, and BF16-convolution/F32-SSM Mamba recurrent state traffic. Compute, router cost, activation traffic, state writes, and prefill traffic are outside Bounds Engine v1.", "notes": "The repo config records bfloat16 weights with tiny F32 correction tensors and mamba_ssm_cache_dtype float32. The model card's vLLM command does not request attention KV-cache quantization, so this profile uses two-byte attention KV plus F32 SSM state." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano 30B BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA cbd3fa9f933d55ef16a84236559f4ee2a0526848, the HF API records a public transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, 1064387 downloads, region:us, and safetensors parameters BF16 31577934400 and F32 5888. The model card states a hybrid MoE architecture with 23 Mamba-2 and MoE layers, 6 attention layers, 128 experts plus 1 shared expert per MoE layer, 6 experts activated per token, 3.5B active parameters, 30B total parameters, default 256K context, optional 1M context, and a vLLM command with --max-model-len 262144." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 52 layers, hybrid_override_pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME, 23 Mamba layers, 23 MoE layers, 6 attention layers, hidden_size 2688, intermediate_size 1856, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 128 routed experts, 6 experts per token, 1 shared expert, mamba_num_heads 64, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, and tie_word_embeddings false." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/raw/97ab8012882a655dc38df4fee47422aca9caca07/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited architecture fields between the base and instruct configs: 52-layer hybrid pattern, hidden size, intermediate sizes, attention geometry, MoE expert counts, Mamba geometry, 262144 max positions, vocabulary size, and untied embeddings. The base config omits torch_dtype and mamba_ssm_cache_dtype fields while preserving the bounds geometry." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. HybridMambaAttentionDynamicCache stores attention key/value lists plus Mamba conv_states and ssm_states. For this repo the mixer convolution spans the 6144-wide hidden+B+C projection and the SSM state shape is batch x 4096 x 128 in each Mamba layer; mamba_ssm_cache_dtype is float32. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads. The MoE path applies routed experts plus an always-on shared expert residual." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/cbd3fa9f933d55ef16a84236559f4ee2a0526848/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 13 safetensors shard headers were range-read directly. The headers match the index tensor count of 6243 and index total size 63.155886464 GB; header payload offsets sum to 63.155892352 GB including 23.552 KB of F32 tensors. Stored bytes group as BF16 63.155868800 GB and F32 0.000023552 GB. The resident-only input embedding is backbone.embeddings.weight, totaling 0.704643072 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 3.701633152 GB. Routed expert tensors total 58.749616128 GB, exactly 0.458981376 GB per expert index across all 23 MoE layers. Mamba conv1d.weight headers have shape [6144, 1, 4], confirming the convolution-state width used with the F32 SSM cache." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, the model card, pinned served config, base config comparison, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, BF16 KV, recurrent state shape, input embedding exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-nano-30b-a3b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8", "title": "NVIDIA Nemotron 3 Nano 30B A3B FP8", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's FP8 Nemotron 3 Nano 30B A3B hybrid Mamba2-Transformer MoE serving artifact.", "model_family": "nemotron-h-nano-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "relation": "quantized", "source": "Hugging Face base_model metadata plus direct served config comparison", "config_compatible": true, "notes": "The FP8 repo declares the BF16 repo as its quantized base. Manual comparison found matching audited architecture fields between the pinned FP8 config and the audited BF16 instruct config; only the transformers_version field differs." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.679949312, "main_resident_weight_gb": 31.97530624, "auxiliary_resident_weight_gb": 0.704643072, "fixed_weight_gb": 2.600451072, "routed_expert_weight_gb": 0.229491056, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_fp8_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding; fixed traffic includes Mamba, attention, router, shared expert, norm, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 9 indexed payload shards and 18179 tensors totaling 32.679949312 GB: F8_E4M3 30.491811840 GB, BF16 2.156424064 GB, and F32 0.031713408 GB. The repo also includes a 40-byte metadata-only tenth safetensors shard that is not referenced by the index and carries no tensor payload. Routed expert tensors are byte-uniform across all 128 expert indexes and 23 MoE layers at exactly 0.229491056 GB per expert index. Non-routed ordinary decode traffic totals 2.600451072 GB, while the resident-only input embedding totals 0.704643072 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 5, 12, 19, 26, 33, and 42. The ModelOpt quantization config and model-card vLLM command use FP8 KV cache, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.049364992, "read_gb_per_output_token": 0.049364992, "state_formula": "23 Mamba layers * (((4096 intermediate scalars + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel) * 2 BF16 bytes + (4096 intermediate scalars * 128 SSM state) * 4 F32 bytes)", "notes": "The served config records mamba_ssm_cache_dtype float32. The custom Mamba mixer derives a 6144-wide convolution state from the 4096 intermediate stream plus grouped B/C SSM projections, and the decode path promotes the SSM recurrent update through float32 A/dA terms. This profile charges BF16 convolution state plus F32 SSM recurrent state, one full fixed Mamba state read per generated token. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for six attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for twenty-three Mamba layers." }, "notes": "NemotronHForCausalLM uses a 52-layer hybrid pattern with 23 Mamba layers, 23 MoE layers, and 6 full-attention layers. The model card advertises support up to 1M context, while the served config records 262144 max positions and the default vLLM command uses 262144; this profile uses the served config value." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp8-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-fp8-hybrid-mamba-moe-fp8-kv", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, FP8 attention KV bytes, and the standalone BF16/F32 Mamba recurrent state. Dequantization, activation traffic, router compute, expert compute, state writes, and prefill traffic are outside Bounds Engine v1.", "notes": "The ModelOpt recipe stores most tensors as FP8, keeps the attention layers and the Mamba layers feeding those attention layers in BF16, stores scale side tensors in F32, and quantizes the KV cache to FP8. The model card's vLLM command uses VLLM_USE_FLASHINFER_MOE_FP8=1 and --kv-cache-dtype fp8." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano 30B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA f8dc1c0afee92f44417695b4f5ddca9afc95ea58, the HF API records a public non-gated transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, base_model nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with base_model:quantized metadata, region:us, 332474 downloads, and safetensors parameter groups F32 7922384, BF16 1078212032, and F8_E4M3 30491811840. The model card states a hybrid MoE architecture with 23 Mamba-2 and MoE layers, 6 attention layers, 128 routed experts plus 1 shared expert per MoE layer, 6 experts activated per token, 3.5B active parameters, 30B total parameters, FP8 weights with FP8 KV cache, default 256K context, optional 1M context, and a vLLM command with --max-model-len 262144 and --kv-cache-dtype fp8." }, { "label": "NVIDIA Nemotron 3 Nano 30B FP8 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/raw/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 52 layers, hybrid_override_pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME, 23 Mamba layers, 23 MoE layers, 6 attention layers, hidden_size 2688, intermediate_size 1856, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 128 routed experts, 6 experts per token, 1 shared expert, mamba_num_heads 64, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, vocab_size 131072, and tie_word_embeddings false." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching audited geometry fields between the BF16 base and FP8 configs: architecture, model type, dtype, 52-layer hybrid pattern, hidden size, intermediate sizes, attention geometry, MoE expert counts, Mamba geometry, mamba_ssm_cache_dtype float32, 262144 max positions, vocabulary size, and untied embeddings." }, { "label": "NVIDIA Nemotron 3 Nano 30B FP8 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/raw/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 0.29.0, quant_algo FP8, kv_cache_quant_algo FP8, group_size 16, and explicit module exclusions for lm_head, the six attention q/k/v/o projections, the Mamba in/out projections feeding attention layers, and all Mamba conv1d modules." }, { "label": "NVIDIA Nemotron 3 Nano 30B FP8 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/resolve/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "All nine indexed safetensors shard headers were range-read directly. Stored tensor bytes match index total_size at 32.679949312 GB across 18179 tensors: F8_E4M3 30.491811840 GB, BF16 2.156424064 GB, and F32 0.031713408 GB. The resident-only input embedding is backbone.embeddings.weight, totaling 0.704643072 GB. Ordinary text resident tensors excluding input embedding sum to 31.975306240 GB. Fixed ordinary text traffic excluding input embedding and routed expert tensors totals 2.600451072 GB. Routed expert tensors total 29.374855168 GB, exactly 0.229491056 GB per expert index across all 23 MoE layers." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/raw/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. HybridMambaAttentionDynamicCache stores attention key/value lists plus Mamba conv_states and ssm_states. The Mamba mixer derives conv_dim as intermediate_size plus two grouped B/C SSM projections, yielding 6144 convolution-state scalars per Mamba layer; the SSM state uses 4096 intermediate scalars by 128 SSM state. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads. The MoE path applies routed experts plus an always-on shared expert residual." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served config, BF16 config comparison, ModelOpt quantization config, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed FP8/BF16 weight traffic, FP8 KV, recurrent state shape, input embedding exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-nano-30b-a3b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4", "title": "NVIDIA Nemotron 3 Nano 30B A3B NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's NVFP4 Nemotron 3 Nano 30B A3B hybrid Mamba2-Transformer MoE serving artifact.", "model_family": "nemotron-h-nano-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "relation": "quantized", "source": "Hugging Face base_model metadata plus direct served config comparison", "config_compatible": true, "notes": "The NVFP4 repo declares the BF16 repo as its quantized base. Manual comparison found matching audited architecture fields between the pinned NVFP4 config and the audited BF16 instruct config; only the transformers_version field differs." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 19.339781632, "main_resident_weight_gb": 18.63513856, "auxiliary_resident_weight_gb": 0.704643072, "fixed_weight_gb": 2.11176192, "routed_expert_weight_gb": 0.12908888, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding; fixed traffic includes Mamba, attention, router, shared expert, norm, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 5 shards and 24147 tensors totaling 19.339781632 GB: U8 15.245905920 GB, F8_E4M3 1.905738240 GB, BF16 2.156424064 GB, and F32 0.031713408 GB. Routed expert tensors are byte-uniform across all 128 expert indexes and 23 MoE layers at exactly 0.129088880 GB per expert index. Non-routed ordinary decode traffic totals 2.111761920 GB, while the resident-only input embedding totals 0.704643072 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 5, 12, 19, 26, 33, and 42. The ModelOpt quantization config and model-card vLLM command use FP8 KV cache, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.049364992, "read_gb_per_output_token": 0.049364992, "state_formula": "23 Mamba layers * (((4096 intermediate scalars + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel) * 2 BF16 bytes + (4096 intermediate scalars * 128 SSM state) * 4 F32 bytes)", "notes": "The served config records mamba_ssm_cache_dtype float32. The custom Mamba mixer derives a 6144-wide convolution state from the 4096 intermediate stream plus grouped B/C SSM projections, and the decode path promotes the SSM recurrent update through float32 A/dA terms. This profile charges BF16 convolution state plus F32 SSM recurrent state, one full fixed Mamba state read per generated token. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for six attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for twenty-three Mamba layers." }, "notes": "NemotronHForCausalLM uses a 52-layer hybrid pattern with 23 Mamba layers, 23 MoE layers, and 6 full-attention layers. The model card advertises support up to 1M context, while the served config records 262144 max positions and the default vLLM command uses 262144; this profile uses the served config value." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-hybrid-mamba-moe-fp8-kv", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, FP8 attention KV bytes, and the standalone BF16/F32 Mamba recurrent state. Dequantization, activation traffic, router compute, expert compute, state writes, and prefill traffic are outside Bounds Engine v1.", "notes": "The ModelOpt recipe stores routed expert projections mostly as packed NVFP4 U8 tensors with F8_E4M3 scale tensors, non-routed quantized tensors as FP8, selected attention/Mamba/lm_head tensors as BF16, and side metadata as F32. The model card's vLLM command uses --kv-cache-dtype fp8." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano 30B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA ce1b118ae66ec705d02c241525192832eb045fd3, the HF API records a public non-gated transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, base_model nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with base_model:quantized metadata, ModelOpt/NVFP4 tags, region:us, 1094976 downloads, and safetensors parameter groups F32 7916416, BF16 1078212032, F8_E4M3 1905738240, and U8 15245905920. The model card states a hybrid MoE architecture with 23 Mamba-2 and MoE layers, 6 attention layers, 128 routed experts plus 1 shared expert per MoE layer, 6 experts activated per token, 3.5B active parameters, 30B total parameters, NVFP4 weights with FP8 KV cache, default 256K context, optional 1M context, and a vLLM command with --max-model-len 262144 and --kv-cache-dtype fp8." }, { "label": "NVIDIA Nemotron 3 Nano 30B NVFP4 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 52 layers, hybrid_override_pattern MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME, 23 Mamba layers, 23 MoE layers, 6 attention layers, hidden_size 2688, intermediate_size 1856, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 128 routed experts, 6 experts per token, 1 shared expert, mamba_num_heads 64, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, vocab_size 131072, and tie_word_embeddings false." }, { "label": "NVIDIA Nemotron 3 Nano 30B BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found matching audited geometry fields between the BF16 base and NVFP4 configs: architecture, model type, dtype, 52-layer hybrid pattern, hidden size, intermediate sizes, attention geometry, MoE expert counts, Mamba geometry, mamba_ssm_cache_dtype float32, 262144 max positions, vocabulary size, and untied embeddings." }, { "label": "NVIDIA Nemotron 3 Nano 30B NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 0.29.0, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and explicit module exclusions for lm_head, the six attention q/k/v/o projections, the Mamba in/out projections feeding attention layers, and all Mamba conv1d modules." }, { "label": "NVIDIA Nemotron 3 Nano 30B NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/resolve/ce1b118ae66ec705d02c241525192832eb045fd3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "All five safetensors shard headers were range-read directly. Stored tensor bytes match index total_size at 19.339781632 GB across 24147 tensors: BF16 2.156424064 GB, F32 0.031713408 GB, F8_E4M3 1.905738240 GB, and U8 15.245905920 GB. The resident-only input embedding is backbone.embeddings.weight, totaling 0.704643072 GB. Ordinary text resident tensors excluding input embedding sum to 18.635138560 GB. Fixed ordinary text traffic excluding input embedding and routed expert tensors totals 2.111761920 GB. Routed expert tensors total 16.523376640 GB, exactly 0.129088880 GB per expert index across all 23 MoE layers." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. HybridMambaAttentionDynamicCache stores attention key/value lists plus Mamba conv_states and ssm_states. The Mamba mixer derives conv_dim as intermediate_size plus two grouped B/C SSM projections, yielding 6144 convolution-state scalars per Mamba layer; the SSM state uses 4096 intermediate scalars by 128 SSM state. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads. The MoE path applies routed experts plus an always-on shared expert residual." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served config, BF16 base config comparison, ModelOpt quantization config, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed NVFP4/FP8/BF16 weight traffic, FP8 KV, recurrent state shape, input embedding exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-nano-4b-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", "title": "NVIDIA Nemotron 3 Nano 4B BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron 3 Nano 4B hybrid Mamba2-Transformer repo.", "model_family": "nemotron-h-nano-hybrid-dense", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-Nano-9B-v2", "relation": "derived_package", "source": "Hugging Face model metadata, model card compression statement, and direct served config comparison", "config_compatible": false, "notes": "The repo metadata records nvidia/NVIDIA-Nemotron-Nano-9B-v2 as the base model, and the model card says this 4B checkpoint was compressed from that parent with Nemotron Elastic. Manual config comparison found different hidden size, layer count, hybrid pattern, context length, chunk size, Mamba head count, and EOS token, so this profile uses the served 4B config directly." }, "architecture": { "canonical_architecture_id": "nemotron-3-nano-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.973556832, "swept_params_b": 3.56251504, "auxiliary_resident_params_b": 0.411041792, "resident_weight_gb": 7.947113664, "swept_weight_gb": 7.12503008, "auxiliary_resident_weight_gb": 0.822083584, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode through backbone.layers, backbone.norm_f, and lm_head.weight, excluding resident-only input embedding", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record one BF16 shard with 263 tensors totaling 3973556832 stored parameters. The checkpoint stores separate backbone.embeddings.weight and lm_head.weight tensors with tie_word_embeddings false; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 4, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 12, 17, 24, and 32. Attention stores K and V with 8 KV heads and 128 head dimension before any query-head repetition." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.042921984, "read_gb_per_output_token": 0.042921984, "state_formula": "21 Mamba layers * (((96 Mamba heads * 80 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) + (96 Mamba heads * 80 Mamba head dim * 128 SSM state)) * 2 BF16 bytes", "notes": "The custom mixer defines Mamba inner width as mamba_num_heads * mamba_head_dim = 7680 and conv_dim as inner width plus B/C state projections = 9728; the safetensors header confirms conv1d.weight shape [9728, 1, 4]. This profile charges one full fixed Mamba conv+SSM state read per generated token; compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for four attention layers plus fixed BF16 Mamba recurrent state for 21 Mamba layers." }, "notes": "NemotronHForCausalLM uses a 42-layer hybrid pattern with 21 Mamba2 layers, 17 MLP-only layers, and 4 full-attention layers. The model card advertises context length up to 262K, and the served config records max_position_embeddings 262144." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-or-vllm-bf16-hybrid-mamba-transformer", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the safetensors header records only BF16 tensors. The model card's vLLM examples do not request KV-cache quantization, so this profile keeps attention KV and Mamba state as BF16." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Nano 4B BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format", "architecture", "max_context_tokens" ], "notes": "At repo SHA dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f, the HF API records a public transformers text-generation repo with custom code, base_model nvidia/NVIDIA-Nemotron-Nano-9B-v2, NVIDIA Nemotron Open Model License metadata, and safetensors parameters BF16: 3973556832. The model card describes a Mamba2-Transformer hybrid compressed from the 9B v2 parent, states 3.97 x 10^9 model parameters, says the architecture is primarily Mamba-2 and MLP layers with four Attention layers, and records context length up to 262K." }, { "label": "NVIDIA Nemotron 3 Nano 4B BF16 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/raw/dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "layer_pattern", "kv_heads", "head_dim", "mamba_recurrent_state", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 42 layers, hybrid_override_pattern M-M-M-MM-M-M*-M-M*-M-M-M*-M-M-MM*-MMM-M-M-, 21 Mamba layers, 17 MLP-only layers, 4 attention layers, hidden_size 3136, intermediate_size 12544, 40 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 262144, sliding_window null, tie_word_embeddings false, vocab_size 131072, mamba_num_heads 96, mamba_head_dim 80, n_groups 8, conv_kernel 4, and ssm_state_size 128." }, { "label": "NVIDIA Nemotron Nano 9B v2 parent config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The parent config records the same NemotronH model type but a different architecture: 56 layers, hidden_size 4480, intermediate_size 15680, max_position_embeddings 131072, a different hybrid pattern, mamba_num_heads 128, mamba_head_dim 80, chunk_size 128, and eos_token_id 12. This confirms the 4B compressed repo is not config-compatible with the parent." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/raw/dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to Mamba2, * to attention, and - to MLP. HybridMambaAttentionDynamicCache stores attention key/value tensors with sequence length and Mamba conv_states plus ssm_states as fixed state. The mixer derives Mamba inner width from 96 heads * 80 head dimension and conv_dim from that inner width plus two grouped SSM projections." }, { "label": "NVIDIA Nemotron 3 Nano 4B BF16 safetensors header", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/resolve/dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "mamba_recurrent_state" ], "notes": "The single safetensors shard header was range-read directly. Stored tensor bytes sum to 7.947113664 GB across 263 BF16 tensors, matching 3973556832 parameters. backbone.embeddings.weight has shape [131072, 3136] and contributes 411041792 parameters / 0.822083584 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. backbone.layers plus backbone.norm_f plus lm_head.weight total 3562515040 parameters / 7.12503008 GB. Layer grouping from the hybrid pattern gives Mamba tensors 3.319544256 GB, MLP tensors 2.675089536 GB, attention tensors 0.308306432 GB, final norm 0.000006272 GB, and lm_head 0.822083584 GB. Mamba conv1d.weight tensors have shape [9728, 1, 4], confirming the recurrent-state conv width used in this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, 9B parent config comparison, custom runtime code, direct safetensors header byte grouping, and local scrape row correction." }, "notes": "This self-contained profile deliberately uses current HF safetensors/header evidence instead of the older scraped catalog estimate, which overstated this repo as 8.8882B parameters." }, { "id": "nvidia--nvidia-nemotron-3-super-120b-a12b-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "title": "NVIDIA Nemotron 3 Super 120B A12B BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron 3 Super 120B A12B hybrid Mamba2-Transformer latent MoE repo.", "model_family": "nemotron-h-super-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16", "relation": "finetune", "source": "Hugging Face model metadata, model card lineage, and direct served config comparison", "config_compatible": false, "notes": "The BF16 instruct repo and BF16 base repo share the audited model geometry fields except max_position_embeddings. The base config records 1048576 while this served repo records 262144, so this profile uses the served instruct config directly." }, "architecture": { "canonical_architecture_id": "nemotron-3-super-120b-a12b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 247.22210816, "main_resident_weight_gb": 240.263714816, "auxiliary_resident_weight_gb": 6.958393344, "fixed_weight_gb": 14.777931776, "routed_expert_weight_gb": 0.44040192, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding and top-level MTP tensors; fixed traffic includes Mamba, attention, router and latent projections, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight and top-level mtp tensors are resident in the package but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 50 shards and 42683 tensors totaling 247.222108160 GB: BF16 247.222024192 GB plus F32 0.000083968 GB. Routed expert tensors are byte-uniform across all 512 expert indexes and 40 MoE layers at exactly 0.440401920 GB per expert index. Non-routed ordinary decode traffic totals 14.777931776 GB, while resident-only input embedding and MTP tensors total 6.958393344 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 7, 16, 25, 36, 47, 58, 69, and 78. The model card's vLLM example serves this repo with --kv-cache-dtype fp8, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.17104896, "read_gb_per_output_token": 0.17104896, "state_formula": "40 Mamba layers * (((128 Mamba heads * 64 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) * 2 BF16 bytes + (128 Mamba heads * 64 Mamba head dim * 128 SSM state) * 4 F32 bytes)", "notes": "The config records mamba_ssm_cache_dtype float32, mamba_num_heads 128, mamba_head_dim 64, n_groups 8, conv_kernel 4, and ssm_state_size 128. The custom runtime allocates conv_states and ssm_states for Mamba layers. This profile follows the documented vLLM serving path: BF16 convolution state plus F32 SSM state, charged as one fixed state read per generated token." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for eight attention layers plus fixed Mamba recurrent state for forty Mamba layers." }, "notes": "NemotronHForCausalLM uses an 88-layer hybrid pattern with 40 Mamba layers, 40 MoE layers, and 8 full-attention layers. The model card advertises up to 1M context while noting that the default Hugging Face config is 256K due to VRAM; this profile uses the served config value 262144." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "fp8-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-bf16-hybrid-mamba-moe-fp8-kv", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes and the documented FP8 attention KV plus Mamba recurrent state traffic. Compute, router cost, activation traffic, state writes, and speculative MTP traffic are outside Bounds Engine v1.", "notes": "The repo config records bfloat16 weights with tiny F32 correction tensors. The model card's vLLM command uses --kv-cache-dtype fp8 and --mamba-ssm-cache-dtype float32; this profile follows that serving configuration for text decode." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Super BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA d51eab0d1f979ebc26b546e634a04f450d99158e, the HF API records a public transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, safetensors parameters BF16 123611012096 and F32 20992, and region:us tags. The model card states 120B total parameters, 12B active parameters, LatentMoE with Mamba-2 plus MoE plus Attention and MTP, support up to 1M context, default Hugging Face config of 256K due to VRAM, and vLLM serving with FP8 KV cache and F32 Mamba SSM cache." }, { "label": "NVIDIA Nemotron 3 Super BF16 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records model_type nemotron_h, NemotronHForCausalLM, bfloat16, 88 layers, hybrid_override_pattern MEMEMEM*EMEMEMEM*EMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEM*EMEMEMEME, 40 Mamba layers, 40 MoE layers, 8 attention layers, hidden_size 4096, intermediate_size 2688, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 512 routed experts, 22 experts per token, 1 shared expert, moe_latent_size 1024, mamba_num_heads 128, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, mtp_hybrid_override_pattern *E, and one next-token prediction layer." }, { "label": "NVIDIA Nemotron 3 Super BF16 base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16/raw/46cc6113d364942e7742b0b2afd35b5db5058b29/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited architecture fields between the BF16 base and BF16 instruct repo except max_position_embeddings: the base records 1048576 and the instruct repo records 262144. This prevents treating the two served configs as fully compatible for bounds purposes." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. NemotronHHybridDynamicCache stores attention key/value caches plus Mamba conv_states and ssm_states. The Mamba mixer derives conv state width from the inner width plus grouped SSM B/C projections, and the MoE path applies routed experts plus a shared expert residual." }, { "label": "NVIDIA Nemotron 3 Super BF16 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/resolve/d51eab0d1f979ebc26b546e634a04f450d99158e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 50 safetensors shard headers were range-read directly. The headers match the index tensor count of 42683 and index total size 247.222108160 GB. Stored bytes group as BF16 247.222024192 GB and F32 0.000083968 GB. Auxiliary resident tensors are backbone.embeddings.weight plus top-level mtp tensors, totaling 6.958393344 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 14.777931776 GB. Routed expert tensors total 225.485783040 GB, exactly 0.440401920 GB per expert index across all 40 MoE layers. Mamba conv1d.weight headers confirm conv state width 10240." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the pinned served config, BF16 base config comparison, model card serving notes, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, MTP exclusion, FP8 KV, F32 Mamba SSM state, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-super-120b-a12b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8", "title": "NVIDIA Nemotron 3 Super 120B A12B FP8", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's FP8 Nemotron 3 Super 120B A12B hybrid Mamba2-Transformer latent MoE serving artifact.", "model_family": "nemotron-h-super-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "relation": "quantized", "source": "Hugging Face model card precision lineage plus direct served config comparison", "config_compatible": true, "notes": "The FP8 repo does not publish base_model metadata, but the model card presents the BF16, FP8, and NVFP4 Super artifacts as the same 120B A12B architecture family. Manual comparison found no differences in the audited geometry fields between the pinned FP8 config and the audited BF16 instruct config; the FP8 repo adds ModelOpt FP8 quantization metadata." }, "architecture": { "canonical_architecture_id": "nemotron-3-super-120b-a12b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 128.334890312, "main_resident_weight_gb": 121.37649696, "auxiliary_resident_weight_gb": 6.958393352, "fixed_weight_gb": 8.63327776, "routed_expert_weight_gb": 0.2202016, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_fp8_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding and top-level MTP tensors; fixed traffic includes Mamba, attention, router and latent projections, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight and top-level mtp tensors are resident in the package but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 26 shards and 124941 tensors totaling 128.334890312 GB: F8_E4M3 118.887546880 GB, BF16 9.446930432 GB, and F32 0.000413000 GB. Routed expert tensors are byte-uniform across all 512 expert indexes and 40 MoE layers at exactly 0.220201600 GB per expert index. Non-routed ordinary decode traffic totals 8.633277760 GB, while resident-only input embedding and MTP tensors total 6.958393352 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 7, 16, 25, 36, 47, 58, 69, and 78. The ModelOpt quantization config and model-card vLLM command use FP8 KV cache, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.17104896, "read_gb_per_output_token": 0.17104896, "state_formula": "40 Mamba layers * (((128 Mamba heads * 64 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) * 2 BF16 bytes + (128 Mamba heads * 64 Mamba head dim * 128 SSM state) * 4 F32 bytes)", "notes": "The served config records mamba_ssm_cache_dtype float32, and the model card's vLLM command sets --mamba-ssm-cache-dtype float32. The custom runtime allocates conv_states with width 10240 and ssm_states with 128 heads, 64 head dimension, and 128 SSM state. This profile charges BF16-sized convolution state plus F32 SSM state, one full fixed Mamba state read per generated token. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for eight attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for forty Mamba layers." }, "notes": "NemotronHForCausalLM uses an 88-layer hybrid pattern with 40 Mamba layers, 40 MoE layers, and 8 full-attention layers. The model card advertises up to 1M context, while the served config records 262144 max positions and the default vLLM command uses 262144; this profile uses the served config value." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp8-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-fp8-hybrid-mamba-moe-fp8-kv", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, FP8 attention KV bytes, and the documented BF16/F32 Mamba recurrent state. Dequantization, activation traffic, router compute, expert compute, state writes, and speculative MTP traffic are outside Bounds Engine v1.", "notes": "The ModelOpt recipe stores most tensors as FP8, keeps embeddings, MTP, ignored linear modules, lm_head, and selected side tensors in BF16 or F32, and quantizes the KV cache to FP8. The model card's vLLM command uses --kv-cache-dtype fp8 and --mamba-ssm-cache-dtype float32." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Super FP8 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA 7d7e5797b8a3c7abbab54033b6004e93e8b6bc91, the HF API records a public non-gated transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, ModelOpt, region:us, 265465 downloads, and safetensors parameter groups F32 20992, BF16 4723465216, and F8_E4M3 118887546880. The model card states 120B total parameters, 12B active parameters, LatentMoE with Mamba-2 plus MoE plus Attention and MTP, support up to 1M context, default vLLM serving at 256K context, and vLLM serving with FP8 KV cache plus F32 Mamba SSM cache." }, { "label": "NVIDIA Nemotron 3 Super FP8 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records model_type nemotron_h, NemotronHForCausalLM, 88 layers, hybrid_override_pattern MEMEMEM*EMEMEMEM*EMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEM*EMEMEMEME, 40 Mamba layers, 40 MoE layers, 8 attention layers, hidden_size 4096, intermediate_size 2688, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 512 routed experts, 22 experts per token, 1 shared expert, moe_latent_size 1024, mamba_num_heads 128, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, mtp_hybrid_override_pattern *E, one next-token prediction layer, quant_method modelopt, FP8 quantization, and FP8 KV cache scheme." }, { "label": "NVIDIA Nemotron 3 Super FP8 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 0.41.0, quant_algo FP8, kv_cache_quant_algo FP8, and exclude_modules entries for Mamba conv1d modules, MoE latent projections, all attention layers, lm_head, and MTP modules. The served config mirrors this with an FP8 Linear quantization group, FP8 KV cache scheme, and 130 ignored module patterns." }, { "label": "NVIDIA Nemotron 3 Super BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited geometry fields between the BF16 instruct config and this FP8 config: architecture, layer pattern, attention geometry, MoE expert counts, Mamba geometry, MTP settings, 262144 max positions, vocabulary size, and untied embeddings. The FP8 artifact adds ModelOpt quantization metadata while preserving the same served architecture." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. NemotronHHybridDynamicCache stores attention key/value caches plus Mamba conv_states and ssm_states. The cache shape is batch x 10240 x 4 for convolution state and batch x 128 x 64 x 128 for SSM state in each Mamba layer. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads. The MoE path applies routed experts plus an always-on shared expert residual." }, { "label": "NVIDIA Nemotron 3 Super FP8 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/resolve/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 26 safetensors shard headers were range-read directly. The headers match the index tensor count of 124941 and index total size 128.334890312 GB. Stored bytes group as F8_E4M3 118.887546880 GB, BF16 9.446930432 GB, and F32 0.000413000 GB. Auxiliary resident tensors are backbone.embeddings.weight plus top-level mtp tensors, totaling 6.958393352 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 8.633277760 GB. Routed expert tensors total 112.743219200 GB, exactly 0.220201600 GB per expert index across all 40 MoE layers. Mamba conv1d.weight headers confirm conv state width 10240." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the model card, pinned served config, BF16 config comparison, ModelOpt quantization config, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed FP8/BF16 weight traffic, FP8 KV, F32 Mamba SSM state, MTP exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-super-120b-a12b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4", "title": "NVIDIA Nemotron 3 Super 120B A12B NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's NVFP4 Nemotron 3 Super 120B A12B hybrid Mamba2-Transformer latent MoE serving artifact.", "model_family": "nemotron-h-super-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "relation": "quantized", "source": "Hugging Face model card precision lineage plus direct served config comparison", "config_compatible": true, "notes": "The NVFP4 repo does not publish base_model metadata, but the model card presents the BF16, FP8, and NVFP4 Super artifacts as the same 120B A12B architecture family. Manual comparison found no differences in the audited geometry fields between the pinned NVFP4 config and the audited BF16 instruct config; the NVFP4 repo adds ModelOpt mixed-precision quantization metadata." }, "architecture": { "canonical_architecture_id": "nemotron-3-super-120b-a12b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 80.297329824, "main_resident_weight_gb": 73.33893648, "auxiliary_resident_weight_gb": 6.958393344, "fixed_weight_gb": 9.92073232, "routed_expert_weight_gb": 0.12386368, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding and top-level MTP tensors; fixed traffic includes Mamba, attention, router and latent projections, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight and top-level mtp tensors are resident in the package but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 17 shards and 165860 tensors totaling 80.297329824 GB: U8 56.382455808 GB, F8_E4M3 11.873353728 GB, BF16 12.041107456 GB, and F32 0.000412832 GB. Routed expert tensors are byte-uniform across all 512 expert indexes and 40 MoE layers at exactly 0.123863680 GB per expert index. Non-routed ordinary decode traffic totals 9.920732320 GB, while resident-only input embedding and MTP tensors total 6.958393344 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 7, 16, 25, 36, 47, 58, 69, and 78. The ModelOpt quantization config and model-card vLLM commands use FP8 KV cache, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.08716288, "read_gb_per_output_token": 0.08716288, "state_formula": "40 Mamba layers * (((128 Mamba heads * 64 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) * 2 bytes + (128 Mamba heads * 64 Mamba head dim * 128 SSM state) * 2 bytes)", "notes": "The custom runtime allocates conv_states with width 10240 and ssm_states with 128 heads, 64 head dimension, and 128 SSM state. The NVFP4 vLLM and DGX Spark serving examples set mamba SSM cache dtype to float16. This profile charges two bytes per scalar for both the convolution state and the SSM state; compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for eight attention layers plus fixed two-byte Mamba recurrent state for forty Mamba layers." }, "notes": "NemotronHForCausalLM uses an 88-layer hybrid pattern with 40 Mamba layers, 40 MoE layers, and 8 full-attention layers. The model card advertises up to 1M context, while the served config records 262144 max positions and the default vLLM command uses 262144; this profile uses the served config value." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8-attention-bf16-conv-fp16-ssm", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-fp16-ssm", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-hybrid-mamba-moe-fp8-kv-fp16-mamba-ssm", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, FP8 attention KV bytes, and the documented float16-sized Mamba recurrent state. Dequantization, activation traffic, router compute, expert compute, state writes, and speculative MTP traffic are outside Bounds Engine v1.", "notes": "The ModelOpt recipe is mixed precision: routed expert projections use NVFP4, non-routed projections and shared experts use FP8, embeddings/MTP/lm_head/selected tensors stay BF16, and F32 scale side tensors are stored. The model card's vLLM command uses --kv-cache-dtype fp8 and --mamba-ssm-cache-dtype float16." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Super NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA 4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6, the HF API records a public transformers text-generation repo with custom code, NVIDIA Nemotron Open Model License metadata, ModelOpt and 8-bit tags, region:us, 946756 downloads, and safetensors parameters BF16 6020553728, F8_E4M3 11873353728, U8 56382455808, and F32 20992. The model card states 120B total parameters, 12B active parameters, LatentMoE with Mamba-2 plus MoE plus Attention and MTP, support up to 1M context, default vLLM serving at 256K context, and vLLM serving with FP8 KV cache plus float16 Mamba SSM cache." }, { "label": "NVIDIA Nemotron 3 Super NVFP4 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records model_type nemotron_h, NemotronHForCausalLM, bfloat16 dtype, 88 layers, hybrid_override_pattern MEMEMEM*EMEMEMEM*EMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEMEM*EMEMEMEM*EMEMEMEME, 40 Mamba layers, 40 MoE layers, 8 attention layers, hidden_size 4096, intermediate_size 2688, 32 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 512 routed experts, 22 experts per token, 1 shared expert, moe_latent_size 1024, mamba_num_heads 128, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mtp_hybrid_override_pattern *E, one next-token prediction layer, quant_method modelopt, mixed-precision quantization, and FP8 KV cache scheme." }, { "label": "NVIDIA Nemotron 3 Super NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 0.43.0.dev63, quant_algo MIXED_PRECISION, kv_cache_quant_algo FP8, 139 FP8 quantized mixer/shared-expert entries, and 40961 NVFP4 quantized entries, almost entirely routed expert projections with group size 16." }, { "label": "NVIDIA Nemotron 3 Super BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited geometry fields between the BF16 instruct config and this NVFP4 config. The NVFP4 artifact adds ModelOpt mixed-precision quantization metadata while preserving the same served architecture and 262144 max position setting." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to mamba, E to moe, and * to attention. NemotronHHybridDynamicCache stores attention key/value caches plus Mamba conv_states and ssm_states. The cache shape is batch x 10240 x 4 for convolution state and batch x 128 x 64 x 128 for SSM state in each Mamba layer. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads." }, { "label": "NVIDIA Nemotron 3 Super NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/resolve/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 17 safetensors shard headers were range-read directly. The headers match the index tensor count of 165860 and index total size 80.297329824 GB. Stored bytes group as U8 56.382455808 GB, F8_E4M3 11.873353728 GB, BF16 12.041107456 GB, and F32 0.000412832 GB. Auxiliary resident tensors are backbone.embeddings.weight plus top-level mtp tensors, totaling 6.958393344 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 9.920732320 GB. Routed expert tensors total 63.418204160 GB, exactly 0.123863680 GB per expert index across all 40 MoE layers. Mamba conv1d.weight headers confirm conv state width 10240." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, the pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed NVFP4/FP8 weight traffic, FP8 KV, float16 Mamba SSM state, MTP exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-ultra-550b-a55b-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "title": "NVIDIA Nemotron 3 Ultra 550B A55B BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron 3 Ultra 550B A55B hybrid Mamba2-Transformer latent MoE repo.", "model_family": "nemotron-h-ultra-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16", "relation": "finetune", "source": "Hugging Face model metadata, model card lineage, and direct served config comparison", "config_compatible": true, "notes": "The BF16 instruct repo and BF16 base repo share the audited model geometry fields: architecture, layer pattern, attention geometry, MoE expert counts, Mamba geometry, MTP settings, 262144 max positions, vocabulary size, and untied embeddings. This profile uses the pinned BF16 instruct config directly for serving bounds." }, "architecture": { "canonical_architecture_id": "nemotron-3-ultra-550b-a55b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1121.049257984, "main_resident_weight_gb": 1096.470552576, "auxiliary_resident_weight_gb": 24.578705408, "fixed_weight_gb": 65.678401536, "routed_expert_weight_gb": 2.01326592, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding and top-level MTP tensors; fixed traffic includes Mamba, attention, router and latent projections, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight and top-level mtp tensors are resident in the package but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 225 shards and 51023 tensors totaling 1121.049257984 GB: BF16 1121.049157632 GB and F32 0.000100352 GB. Routed expert tensors are byte-uniform across all 512 expert indexes and 48 MoE layers at exactly 2.013265920 GB per expert index. Non-routed ordinary decode traffic totals 65.678401536 GB, while resident-only input embedding and MTP tensors total 24.578705408 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The layers_block_type array maps attention layers to indexes 7, 14, 23, 32, 39, 48, 57, 64, 73, 82, 89, and 98. The model-card vLLM/SGLang serving guidance includes FP8 KV cache for production serving, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.20840448, "read_gb_per_output_token": 0.20840448, "state_formula": "48 Mamba layers * (((256 Mamba heads * 64 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) * 2 bytes + (256 Mamba heads * 64 Mamba head dim * 128 SSM state) * 2 bytes)", "notes": "The config records mamba_ssm_cache_dtype float32, but the model-card vLLM commands set --mamba-ssm-cache-dtype float16. The Nemotron-H runtime allocates conv_states with width mamba_num_heads * mamba_head_dim + 2 * n_groups * ssm_state_size and SSM state shape batch x mamba_num_heads x mamba_head_dim x ssm_state_size. This profile charges two bytes per scalar for both convolution state and SSM state; compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for twelve attention layers plus fixed two-byte Mamba recurrent state for forty-eight Mamba layers." }, "notes": "NemotronHForCausalLM uses a 108-layer hybrid pattern with 48 Mamba layers, 48 MoE layers, and 12 full-attention layers. The model card advertises up to 1M context, while the served config records 262144 max positions and vLLM examples use 262144; this profile uses the served config value." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "fp8-attention-fp16-mamba-state", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-fp16-mamba-state", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-bf16-hybrid-mamba-moe-fp8-kv-fp16-mamba-ssm", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes, FP8 attention KV bytes, and the documented float16-sized Mamba recurrent state. Compute, router cost, activation traffic, state writes, and speculative MTP traffic are outside Bounds Engine v1.", "notes": "The repo stores BF16 weights with tiny F32 side tensors. The model card's vLLM/SGLang serving guidance uses FP8 KV cache and sets --mamba-ssm-cache-dtype float16 for vLLM; this profile follows that practical serving configuration for ordinary text decode." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Ultra BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA 624ba927cfbef0427354998700de3d51173c8c04, the HF API records a public transformers text-generation repo with Nemotron-H, latent-moe, MTP, endpoints_compatible, region:us, 110467 downloads, license other/openmdw-1.1, and safetensors parameters BF16 560524578816 and F32 25088. The model card states 550B total parameters, 55B active parameters, LatentMoE with Mamba-2 plus MoE plus Attention and MTP, support up to 1M context, recommended vLLM serving at 262144 context, FP8 KV cache in the multi-node command, and float16 Mamba SSM cache." }, { "label": "NVIDIA Nemotron 3 Ultra BF16 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16/raw/624ba927cfbef0427354998700de3d51173c8c04/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records model_type nemotron_h, NemotronHForCausalLM, bfloat16, 108 layers, 48 Mamba layers, 48 MoE layers, 12 attention layers, hidden_size 8192, intermediate_size 5120, 64 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 512 routed experts, 22 experts per token, 1 shared expert, moe_latent_size 2048, mamba_num_heads 256, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mamba_ssm_cache_dtype float32, mtp_layers_block_type attention plus moe, and one next-token prediction layer." }, { "label": "NVIDIA Nemotron 3 Ultra BF16 base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16/raw/1493b9c9ca301d6b6e5b0bee90fef1990ba49ed4/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences across audited geometry fields between the BF16 base config and the BF16 instruct config: architecture, model type, hidden/intermediate sizes, layer pattern, attention geometry, MoE expert counts, Mamba geometry, MTP settings, 262144 max positions, vocabulary size, and untied embeddings." }, { "label": "Nemotron-H runtime cache implementation reference", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "The Ultra repo is packaged for native transformers Nemotron-H support and does not ship modeling_nemotron_h.py. Manual review of the pinned Nemotron-H runtime implementation in the sibling Super NVFP4 package found NemotronHHybridDynamicCache storing attention key/value caches plus Mamba conv_states and ssm_states. Applying Ultra's pinned config gives conv width 18432 and SSM state shape batch x 256 x 64 x 128 for each Mamba layer. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads." }, { "label": "NVIDIA Nemotron 3 Ultra BF16 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16/resolve/624ba927cfbef0427354998700de3d51173c8c04/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 225 safetensors shard headers were range-read directly. The headers match the index tensor count of 51023 and index total size 1121.049257984 GB. Stored bytes group as BF16 1121.049157632 GB and F32 0.000100352 GB. Auxiliary resident tensors are backbone.embeddings.weight plus top-level mtp tensors, totaling 24.578705408 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 65.678401536 GB. Routed expert tensors total 1030.792151040 GB, exactly 2.013265920 GB per expert index across all 48 MoE layers. Mamba conv1d.weight headers confirm conv state width 18432." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, the model card, pinned served config, BF16 base config comparison, Nemotron-H runtime cache implementation review, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, MTP exclusion, FP8 KV, float16 Mamba SSM state, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-3-ultra-550b-a55b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4", "title": "NVIDIA Nemotron 3 Ultra 550B A55B NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's NVFP4 Nemotron 3 Ultra 550B A55B hybrid Mamba2-Transformer latent MoE serving artifact.", "model_family": "nemotron-h-ultra-hybrid-moe", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16", "relation": "derived_package", "source": "Hugging Face model card architecture lineage plus direct served config comparison", "config_compatible": true, "notes": "The model card identifies the public BF16 Base repo as the Ultra 550B A55B pretraining base and presents the NVFP4 artifact as the same Nemotron-H Ultra architecture family. Manual comparison found no differences in the audited geometry fields between the pinned NVFP4 config and the pinned BF16 Base config; the NVFP4 repo adds ModelOpt mixed-precision quantization metadata and serving/chat packaging." }, "architecture": { "canonical_architecture_id": "nemotron-3-ultra-550b-a55b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 352.28406128, "main_resident_weight_gb": 327.705355872, "auxiliary_resident_weight_gb": 24.578705408, "fixed_weight_gb": 37.794670176, "routed_expert_weight_gb": 0.566231808, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_modelopt_mixed_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through backbone layers and lm_head excluding resident-only input embedding and top-level MTP tensors; fixed traffic includes Mamba, attention, router and latent projections, shared experts, norms, and lm_head tensors", "auxiliary_scope": "backbone.embeddings.weight and top-level mtp tensors are resident in the package but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records one shared expert per MoE layer. Because shared experts are always active, their traffic is folded into fixed_weight_gb rather than charged again as routed expert traffic.", "notes": "Range-read safetensors headers record 113 shards and 198887 tensors totaling 352.284061280 GB: U8 257.698037760 GB, F8_E4M3 60.498640896 GB, BF16 33.281581056 GB, and F32 0.805801568 GB. Routed expert tensors are byte-uniform across all 512 expert indexes and 48 MoE layers at exactly 0.566231808 GB per expert index. Non-routed ordinary decode traffic totals 37.794670176 GB, while resident-only input embedding and MTP tensors total 24.578705408 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The layers_block_type array maps attention layers to indexes 7, 14, 23, 32, 39, 48, 57, 64, 73, 82, 89, and 98. The ModelOpt quantization config and model-card vLLM commands use FP8 KV cache, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.20840448, "read_gb_per_output_token": 0.20840448, "state_formula": "48 Mamba layers * (((256 Mamba heads * 64 Mamba head dim + 2 * 8 groups * 128 SSM state) * 4 convolution kernel) * 2 bytes + (256 Mamba heads * 64 Mamba head dim * 128 SSM state) * 2 bytes)", "notes": "The Nemotron-H runtime allocates conv_states with width mamba_num_heads * mamba_head_dim + 2 * n_groups * ssm_state_size and ssm_states with mamba_num_heads, mamba_head_dim, and ssm_state_size dimensions. The NVFP4 vLLM and SGLang serving examples set mamba SSM cache dtype to float16 even though the config default records float32. This profile charges two bytes per scalar for both convolution state and SSM state; compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as FP8 full-context attention K/V for twelve attention layers plus fixed two-byte Mamba recurrent state for forty-eight Mamba layers." }, "notes": "NemotronHForCausalLM uses a 108-layer hybrid pattern with 48 Mamba layers, 48 MoE layers, and 12 full-attention layers. The model card advertises up to 1M context, while the served config records 262144 max positions and default vLLM/SGLang commands use 262144; this profile uses the served config value." }, "serving": { "weight_format": "fp4_fp8_mixed", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8-attention-fp16-mamba-state", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-fp16-mamba-state", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-hybrid-mamba-moe-fp8-kv-fp16-mamba-ssm", "dequantization_notes": "The memory-side bound charges exact stored ModelOpt tensor bytes, FP8 attention KV bytes, and the documented float16-sized Mamba recurrent state. Dequantization, activation traffic, router compute, expert compute, state writes, and speculative MTP traffic are outside Bounds Engine v1.", "notes": "The ModelOpt recipe is mixed precision: routed expert projections use NVFP4, non-routed projections and shared experts use FP8, embeddings/MTP/lm_head/selected tensors stay BF16 or F32, and U8 side tensors are stored. The model card's vLLM command uses --kv-cache-dtype fp8 and --mamba-ssm-cache-dtype float16." }, "evidence": [ { "label": "NVIDIA Nemotron 3 Ultra NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "active_params_b", "model_family", "max_context_tokens", "serving" ], "notes": "At repo SHA 183968f87ae4cedce3039313cac1fd43d112c578, the HF API records a public transformers text-generation repo with Nemotron-H, latent-moe, MTP, ModelOpt, 8-bit, endpoints_compatible, region:us, 489280 downloads, and safetensors parameters F32 201351896, BF16 16640790528, F8_E4M3 60498640896, U8 257698037760, and logical total 335038821080. The model card states 550B total parameters, 55B active parameters, LatentMoE with Mamba-2 plus MoE plus Attention and MTP, support up to 1M context, default vLLM/SGLang serving at 256K context, and serving with FP8 KV cache plus float16 Mamba SSM cache." }, { "label": "NVIDIA Nemotron 3 Ultra NVFP4 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/raw/183968f87ae4cedce3039313cac1fd43d112c578/config.json", "source_type": "config", "supports": [ "layers", "layer_pattern", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "max_context_tokens", "serving" ], "notes": "The served config records model_type nemotron_h, NemotronHForCausalLM, 108 layers, 48 Mamba layers, 48 MoE layers, 12 attention layers, hidden_size 8192, intermediate_size 5120, 64 attention heads, 2 KV heads, head_dim 128, max_position_embeddings 262144, 512 routed experts, 22 experts per token, 1 shared expert, moe_latent_size 2048, mamba_num_heads 256, mamba_head_dim 64, n_groups 8, conv_kernel 4, ssm_state_size 128, mtp_layers_block_type attention plus moe, one next-token prediction layer, quant_method modelopt, mixed-precision quantization, and FP8 KV cache scheme." }, { "label": "NVIDIA Nemotron 3 Ultra NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/raw/183968f87ae4cedce3039313cac1fd43d112c578/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "routed_expert_weight_format" ], "notes": "hf_quant_config.json records ModelOpt 1.0.0, quant_algo MIXED_PRECISION, kv_cache_quant_algo FP8, 192 FP8 quantized mixer/shared-expert entries, and 49152 NVFP4 quantized routed expert projection entries with group size 16." }, { "label": "NVIDIA Nemotron 3 Ultra BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16/raw/1493b9c9ca301d6b6e5b0bee90fef1990ba49ed4/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences across 25 audited geometry fields between the BF16 Base config and this NVFP4 config: architecture, model type, hidden/intermediate sizes, layer pattern, attention geometry, MoE expert counts, Mamba geometry, MTP layer pattern, 262144 max positions, vocabulary size, and untied embeddings. The NVFP4 artifact adds ModelOpt mixed-precision quantization metadata while preserving the same served architecture." }, { "label": "Nemotron-H runtime cache implementation reference", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "moe_weight_traffic" ], "notes": "The Ultra repo is packaged for native transformers Nemotron-H support and does not ship modeling_nemotron_h.py. Manual review of the pinned Nemotron-H runtime implementation in the sibling Super NVFP4 package found NemotronHHybridDynamicCache storing attention key/value caches plus Mamba conv_states and ssm_states. The implementation derives conv state width as mamba_num_heads * mamba_head_dim + 2 * n_groups * ssm_state_size and SSM state shape as batch x mamba_num_heads x mamba_head_dim x ssm_state_size. Applying Ultra's pinned config gives conv width 18432 and SSM state shape batch x 256 x 64 x 128 for each Mamba layer. Attention projections store K and V with 2 KV heads and 128 head dimension before repeating to query heads." }, { "label": "NVIDIA Nemotron 3 Ultra NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/resolve/183968f87ae4cedce3039313cac1fd43d112c578/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "mamba_recurrent_state" ], "notes": "All 113 safetensors shard headers were range-read directly. The headers match the index tensor count of 198887 and index total size 352.284061280 GB. Stored bytes group as U8 257.698037760 GB, F8_E4M3 60.498640896 GB, BF16 33.281581056 GB, and F32 0.805801568 GB. Auxiliary resident tensors are backbone.embeddings.weight plus top-level mtp tensors, totaling 24.578705408 GB. Ordinary fixed text traffic excluding input embedding and routed expert tensors totals 37.794670176 GB. Routed expert tensors total 289.910685696 GB, exactly 0.566231808 GB per expert index across all 48 MoE layers. Mamba conv1d.weight headers confirm conv state width 18432." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, Nemotron-H runtime cache implementation review, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "Production code must use this self-contained profile rather than deriving hybrid Mamba/MoE, mixed NVFP4/FP8/BF16/F32 weight traffic, FP8 KV, float16 Mamba SSM state, MTP exclusion, or exact routed-expert traffic from the repo name." }, { "id": "nvidia--nvidia-nemotron-nano-12b-v2-vl-bf16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16", "title": "NVIDIA Nemotron Nano 12B v2 VL BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron Nano 12B v2 vision-language package.", "model_family": "nemotron-h-nano-vl-hybrid-dense", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-Nano-12B-v2", "relation": "derived_package", "source": "Model card language-encoder statement, served VL config, parent text config comparison, and direct safetensors header grouping", "config_compatible": false, "notes": "The model card identifies the language encoder as NVIDIA-Nemotron-Nano-12B-v2 and the vision encoder as C-RADIOv2-H. The embedded llm_config matches the 12B parent text geometry on layer count, hybrid pattern, hidden size, MLP width, attention heads, KV heads, Mamba dimensions, context length, and untied embeddings, but the VL package expands vocab_size to 132096 for image/video tokens and wraps the language model with a vision encoder and projector. This profile therefore treats the served VL config and tensor headers as authoritative." }, "architecture": { "canonical_architecture_id": "nemotron-nano-12b-v2-vl-bf16", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 13.181860358, "swept_params_b": 11.644155392, "auxiliary_resident_params_b": 1.537704966, "resident_weight_gb": 26.363720728, "swept_weight_gb": 23.288310784, "auxiliary_resident_weight_gb": 3.075409944, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "ordinary cached text decode through language_model.backbone.layers, language_model.backbone.norm_f, and language_model.lm_head.weight", "auxiliary_scope": "language_model.backbone.embeddings, vision_model, and mlp1 projector are resident for prompt/image processing but are not swept as full matrices for each generated text token", "notes": "Range-read safetensors headers record seven shards with 761 tensors totaling 13.181860358B stored parameters / 26.363720728 GB. The checkpoint stores separate language_model.backbone.embeddings.weight and language_model.lm_head.weight tensors with tie_word_embeddings false; ordinary cached text decode excludes the input embedding, vision encoder, and projector, but includes the separate lm_head output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 7, 15, 23, 31, 39, and 47. Attention stores K and V with 8 KV heads and 128 head dimension before any query-head repetition." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.149553152, "read_gb_per_output_token": 0.149553152, "state_formula": "28 Mamba layers * (((128 Mamba heads * 80 Mamba head dim + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel * 2 BF16 bytes) + (128 Mamba heads * 80 Mamba head dim * 128 SSM state * 4 F32 bytes))", "notes": "The custom mixer defines Mamba inner width as 128 heads * 80 head dimension = 10240 and conv_dim as that inner width plus two grouped SSM projections = 12288; safetensors headers confirm conv1d.weight shape [12288, 1, 4]. The HybridMambaAttentionDynamicCache stores BF16 convolution state plus F32 SSM state. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for six attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for 28 Mamba layers. The vision encoder and projector run during multimodal prefill, not per generated text token." }, "notes": "NemotronH_Nano_VL_V2 wraps a 62-layer Nemotron H language model with 28 Mamba2 layers, 28 MLP-only layers, and 6 full-attention layers, plus a C-RADIOv2-H vision encoder and two-layer projector. This profile models ordinary cached text decode after any image/video embeddings have been injected into the prompt." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000000910342, "kv_store_format": "bf16-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-vlm-hybrid-mamba-transformer-f32-ssm", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes, BF16 attention KV, BF16 Mamba convolution state, and F32 Mamba SSM state. Vision prefill, projector compute, state writes, and activation traffic are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16. Direct headers record BF16 tensors plus six F32 scalar parameters, giving an effective stored 2.000000000910342 bytes per API parameter." }, "evidence": [ { "label": "NVIDIA Nemotron Nano 12B v2 VL BF16 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "weight_format", "architecture", "max_context_tokens", "serving" ], "notes": "At repo SHA 5d250e2e111dc5e1434131bdf3d590c27a878ade, the live API records a public non-gated image-text-to-text Transformers repo with NVIDIA Open Model License metadata, endpoints_compatible, region:us, 174551 downloads, and safetensors parameters BF16 13181860352, F32 6, total 13181860358. The model card describes a 12.6B vision-language model with C-RADIOv2-H vision encoder, NVIDIA-Nemotron-Nano-12B-v2 language encoder, up to four images, video support, and 128K input+output tokens." }, { "label": "NVIDIA Nemotron Nano 12B v2 VL BF16 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/raw/5d250e2e111dc5e1434131bdf3d590c27a878ade/config.json", "source_type": "config", "supports": [ "model_family", "layers", "layer_pattern", "kv_heads", "head_dim", "mamba_recurrent_state", "vision_encoder", "projector", "max_context_tokens", "serving" ], "notes": "The served config records NemotronH_Nano_VL_V2, BF16 dtype, max_sequence_length 131072, force_image_size 512, patch_size 16, downsample_ratio 0.5, vit_hidden_size 1280, projector_hidden_size 20480, and a nested NemotronHForCausalLM llm_config. The llm_config records 62 layers, hybrid pattern M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M-, 28 Mamba layers, 28 MLP-only layers, 6 attention layers, hidden_size 5120, intermediate_size 20480, 40 attention heads, 8 KV heads, attention_head_dim 128, max_position_embeddings 131072, sliding_window null, tie_word_embeddings false, vocab_size 132096, mamba_num_heads 128, mamba_head_dim 80, n_groups 8, conv_kernel 4, and ssm_state_size 128." }, { "label": "NVIDIA Nemotron Nano 12B v2 parent config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2/raw/f428df0ec725fed457b89cfca54dc26500fb88c1/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the embedded VL llm_config matches the parent text geometry on layer count, hidden size, MLP width, hybrid pattern, attention heads, KV heads, Mamba dimensions, context length, sliding_window null, and untied embeddings. The package differs in vocab_size 132096 vs parent 131072 and records attention_head_dim 128 where the parent uses head_dim 128." }, { "label": "NVIDIA Nemotron Nano 12B v2 VL custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/raw/5d250e2e111dc5e1434131bdf3d590c27a878ade/modeling.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry", "vision_encoder", "projector" ], "notes": "Manual review found NemotronH_Nano_VL_V2 constructing AutoModelForCausalLM from llm_config, AutoModel from vision_config, and mlp1 as RMSNorm plus a 5120->20480->5120 BF16 projector after pixel shuffle. During generation, image/video embeddings are computed before calling language_model.generate with inputs_embeds, so ordinary text decode sweeps the language model while the vision encoder and projector are resident/prefill-side. modeling_nemotron_h.py defines HybridMambaAttentionDynamicCache with attention key/value tensors and Mamba conv_states plus F32 ssm_states." }, { "label": "NVIDIA Nemotron Nano 12B v2 VL BF16 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/5d250e2e111dc5e1434131bdf3d590c27a878ade/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "vision_encoder", "projector", "mamba_recurrent_state" ], "notes": "All seven safetensors shard headers were range-read directly. Stored tensor bytes sum to 26.363720728 GB across 761 tensors, matching index total_size 26363720728 bytes and API total 13181860358 parameters. Payloads split into BF16 26.363720704 GB and F32 0.000000024 GB. language_model.backbone.embeddings.weight contributes 1.352663040 GB resident-only for ordinary decode, and language_model.lm_head.weight is stored separately with the same shape and remains swept. language_model.backbone.layers plus backbone.norm_f plus lm_head total 23.288310784 GB swept traffic. vision_model tensors total 1.303306264 GB and mlp1 projector tensors total 0.419440640 GB, both resident/prefill-side for ordinary decode. Language layer bytes split into Mamba 9.436263424 GB, attention 0.755036160 GB, and MLP-only 11.744337920 GB. Mamba conv1d.weight tensors have shape [12288, 1, 4], confirming the convolution-state width used in this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served VL config, parent 12B text config comparison, custom runtime code, image processor config, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "This profile supersedes the generated dense no-config catalog estimate. It is scoped to ordinary text decode after multimodal prefill; it does not attempt to benchmark image/video prefill, dynamic tile counts, video pruning, or vision encoder throughput." }, { "id": "nvidia--nvidia-nemotron-nano-9b-v2-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8", "title": "NVIDIA Nemotron Nano 9B v2 FP8", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's FP8 Nemotron Nano 9B v2 hybrid Mamba2-Transformer serving artifact.", "model_family": "nemotron-h-nano-hybrid-dense", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-Nano-9B-v2", "relation": "quantized", "source": "Hugging Face model card base_model metadata, model card quantization statement, served config comparison, and ModelOpt quantization config", "config_compatible": true, "notes": "The FP8 model card states this repo is a quantized version of nvidia/NVIDIA-Nemotron-Nano-9B-v2. Manual comparison found no differences across 18 checked architecture fields between the FP8 config and the audited BF16 9B v2 config. The FP8 repo adds ModelOpt quantization sidecar metadata while preserving the 9B v2 geometry." }, "architecture": { "canonical_architecture_id": "nemotron-nano-9b-v2", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.888227536, "swept_params_b": 8.301024976, "auxiliary_resident_params_b": 0.58720256, "resident_weight_gb": 10.285035328, "swept_weight_gb": 9.110630208, "auxiliary_resident_weight_gb": 1.17440512, "resident_parameter_scope": "safetensors_header_modelopt_fp8_bf16_f32_tensor_elements", "swept_parameter_scope": "ordinary text decode through backbone.layers, backbone.norm_f, and lm_head.weight, excluding resident-only input embedding", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record the exact mixed ModelOpt representation rather than an ideal one-byte-per-parameter checkpoint. The config records tie_word_embeddings false, and the checkpoint stores separate BF16 backbone.embeddings.weight and lm_head.weight tensors. Ordinary decode charges lm_head.weight as the output projection and keeps backbone.embeddings.weight resident-only." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 4, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 14, 21, 30, and 39. hf_quant_config.json records kv_cache_quant_algo null, and the model-card vLLM command does not request FP8 KV cache, so attention K/V are charged as BF16 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.144211968, "read_gb_per_output_token": 0.144211968, "state_formula": "27 Mamba layers * (((128 Mamba heads * 80 Mamba head dim + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel * 2 BF16 bytes) + (128 Mamba heads * 80 Mamba head dim * 128 SSM state * 4 F32 bytes))", "notes": "The custom mixer derives Mamba inner width as 128 heads * 80 head dimension = 10240 and conv_dim as that inner width plus B/C state projections = 12288; safetensors headers confirm conv1d.weight shape [12288, 1, 4]. The model-card vLLM command requires --mamba_ssm_cache_dtype float32 for accurate quality, so this profile charges BF16 convolution state plus F32 SSM state. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for four attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for 27 Mamba layers." }, "notes": "NemotronHForCausalLM uses a 56-layer hybrid pattern with 27 Mamba2 layers, 25 MLP-only layers, and 4 full-attention layers. The model card and served config record 128K context." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-modelopt-fp8-hybrid-mamba-transformer-f32-ssm", "dequantization_notes": "The memory-side bound charges exact ModelOpt stored tensor bytes, BF16 attention KV bytes, and the BF16/F32 Mamba recurrent state. FP8 dequantization, activation traffic, state writes, and compute are outside Bounds Engine v1.", "notes": "The ModelOpt quantization sidecar records quant_algo FP8, kv_cache_quant_algo null, group_size 16, and exclusions for lm_head, all attention q/k/v/o projections, all Mamba conv1d modules, and selected Mamba projections. The model card says Mamba and MLP layers are quantized to FP8 while all four attention layers and Conv1d components remain BF16." }, "evidence": [ { "label": "NVIDIA Nemotron Nano 9B v2 FP8 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format", "architecture", "max_context_tokens", "serving" ], "notes": "At repo SHA 8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0, the HF API records a public non-gated transformers text-generation repo with custom Nemotron H code, NVIDIA Open Model License metadata, base_model nvidia/NVIDIA-Nemotron-Nano-9B-v2, region:us, 227137 downloads, and safetensors parameter groups F32 104, BF16 1396807168, and F8_E4M3 7491420160. The model card states this is a quantized version of the BF16 9B v2 model, with 56 layers total: 27 Mamba, 25 MLP, and 4 attention layers." }, { "label": "NVIDIA Nemotron Nano 9B v2 FP8 model card serving notes", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8", "source_type": "model_card", "supports": [ "serving", "weight_format", "mamba_recurrent_state" ], "notes": "The card states that in the FP8 quantized version the Mamba and MLP layers are quantized to FP8, while all four attention layers and Conv1d components within the Mamba layers remain BF16. The vLLM examples run vllm serve with --trust-remote-code, --max-num-seqs 64, and --mamba_ssm_cache_dtype float32, and warn that omitting the float32 Mamba SSM cache option may degrade quality. The vLLM command does not request FP8 KV cache." }, { "label": "NVIDIA Nemotron Nano 9B v2 FP8 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "layer_pattern", "kv_heads", "head_dim", "mamba_recurrent_state", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 56 layers, hybrid_override_pattern M-M-M-MM-M-M-M*-M-M-M*-M-M-M-M*-M-M-M-M*-M-MM-M-M-M-M-M-, 27 Mamba layers, 25 MLP-only layers, 4 attention layers, hidden_size 4480, intermediate_size 15680, 40 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, sliding_window null, tie_word_embeddings false, vocab_size 131072, mamba_num_heads 128, mamba_head_dim 80, n_groups 8, conv_kernel 4, and ssm_state_size 128." }, { "label": "NVIDIA Nemotron Nano 9B v2 BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences across 18 checked architecture fields between the FP8 config and the audited BF16 9B v2 config: architecture, model type, 56-layer hybrid pattern, hidden size, intermediate size, attention geometry, Mamba geometry, max positions, vocabulary size, sliding window, and untied embeddings." }, { "label": "NVIDIA Nemotron Nano 9B v2 FP8 ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "quantized_module_scope" ], "notes": "hf_quant_config.json records ModelOpt 0.34.1.dev141, quant_algo FP8, kv_cache_quant_algo null, group_size 16, and explicit module exclusions for lm_head, all attention q/k/v/o projections, all Mamba conv1d modules, and selected Mamba projection modules." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to Mamba2, * to attention, and - to MLP. HybridMambaAttentionDynamicCache stores attention key/value tensors with sequence length and Mamba conv_states plus ssm_states as fixed state. NemotronHMamba2Mixer derives conv_dim from Mamba inner width plus two grouped SSM projections, and the single-token update path runs causal_conv1d_update over that full hidden+B+C projection." }, { "label": "NVIDIA Nemotron Nano 9B v2 FP8 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/resolve/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "mamba_recurrent_state" ], "notes": "The index references three payload shards and records total_size 10.285035328 GB. Range-reading the indexed safetensors headers found 549 tensors totaling exactly 10.285035328 GB: 7.491420160 GB F8_E4M3, 2.793614336 GB BF16, and 0.000000832 GB F32. Direct header tensor elements total 8.888227536B, 104 scalar F32 elements above the API safetensors total. backbone.embeddings.weight has shape [131072, 4480] and contributes 0.587202560B elements / 1.174405120 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. backbone.layers plus backbone.norm_f plus lm_head.weight total 8.301024976B tensor elements / 9.110630208 GB. Layer grouping from the hybrid pattern gives Mamba layer tensors 3.983233968 GB, MLP layer tensors 3.512544400 GB, attention layer tensors 0.440437760 GB, final norm 0.000008960 GB, and lm_head 1.174405120 GB. Mamba conv1d.weight tensors have shape [12288, 1, 4], confirming the convolution-state width used in this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the model card, pinned served config, audited BF16 config comparison, ModelOpt quantization config, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "This profile supersedes the generated metadata estimate, which treated the repo as a flat FP8 dense transformer, missed the hybrid Mamba/attention state adapter, missed BF16 attention and conv1d exclusions, and overstated ordinary KV cache traffic." }, { "id": "nvidia--nvidia-nemotron-nano-9b-v2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVIDIA-Nemotron-Nano-9B-v2", "title": "NVIDIA Nemotron Nano 9B v2 BF16", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's BF16 Nemotron Nano 9B v2 hybrid Mamba2-Transformer reasoning model.", "model_family": "nemotron-h-nano-hybrid-dense", "base_model_proof": { "base_model": "nvidia/NVIDIA-Nemotron-Nano-12B-v2", "relation": "finetune", "source": "Hugging Face model card base_model metadata and direct served config comparison", "config_compatible": false, "notes": "The model card metadata records NVIDIA-Nemotron-Nano-12B-v2 and NVIDIA-Nemotron-Nano-12B-v2-Base as bases. Manual comparison found different architecture geometry in the served 9B config: fewer layers, smaller hidden size, smaller MLP width, different hybrid pattern, and different Mamba geometry from the 12B parent. This profile therefore uses the served 9B config directly." }, "architecture": { "canonical_architecture_id": "nemotron-nano-9b-v2", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.888227328, "swept_params_b": 8.301024768, "auxiliary_resident_params_b": 0.58720256, "resident_weight_gb": 17.776454656, "swept_weight_gb": 16.602049536, "auxiliary_resident_weight_gb": 1.17440512, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode through backbone.layers, backbone.norm_f, and lm_head.weight, excluding resident-only input embedding", "auxiliary_scope": "backbone.embeddings.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record four BF16 shards with 341 tensors totaling 8888227328 stored parameters. The checkpoint stores separate backbone.embeddings.weight and lm_head.weight tensors with tie_word_embeddings false; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 4, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The hybrid_override_pattern maps '*' to full-attention layers at indexes 14, 21, 30, and 39. Attention stores K and V with 8 KV heads and 128 head dimension before any query-head repetition." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.144211968, "read_gb_per_output_token": 0.144211968, "state_formula": "27 Mamba layers * (((128 Mamba heads * 80 Mamba head dim + 2 * 8 groups * 128 SSM state scalars) * 4 convolution kernel * 2 BF16 bytes) + (128 Mamba heads * 80 Mamba head dim * 128 SSM state * 4 F32 bytes))", "notes": "The custom mixer defines Mamba inner width as 128 heads * 80 head dimension = 10240 and conv_dim as that inner width plus B/C state projections = 12288; safetensors headers confirm conv1d.weight shape [12288, 1, 4]. The model card's vLLM instructions explicitly require --mamba_ssm_cache_dtype float32 for accuracy, so this profile charges BF16 convolution state plus F32 SSM state. Compute, state writes, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid text decode is represented as BF16 full-context K/V for four attention layers plus fixed BF16-convolution/F32-SSM Mamba recurrent state for 27 Mamba layers." }, "notes": "NemotronHForCausalLM uses a 56-layer hybrid pattern with 27 Mamba2 layers, 25 MLP-only layers, and 4 full-attention layers. The model card and served config record 128K context." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16-attention-bf16-conv-f32-ssm", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16-attention-bf16-conv-f32-ssm", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-bf16-hybrid-mamba-transformer-f32-ssm", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. The memory-side bound charges exact stored safetensors bytes, BF16 attention KV, BF16 Mamba convolution state, and F32 Mamba SSM state. Compute, state writes, and activation traffic are outside Bounds Engine v1.", "notes": "The repo config records torch_dtype bfloat16, and safetensors headers record only BF16 tensors. The model card's vLLM examples require --mamba_ssm_cache_dtype float32, so the profile uses F32 SSM state even though the attention KV and convolution state remain BF16." }, "evidence": [ { "label": "NVIDIA Nemotron Nano 9B v2 model card and API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-9B-v2", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format", "architecture", "max_context_tokens", "serving" ], "notes": "At repo SHA 6533e8de2c68e4536bf7c411d7a3ce5734111476, the HF API records a public transformers text-generation repo with custom Nemotron H code, NVIDIA Open Model License metadata, region:us, 728827 downloads, and safetensors parameters BF16: 8888227328. The model card describes NVIDIA-Nemotron-Nano-9B-v2 as a hybrid Mamba-2/MLP model with just four Attention layers, records context length up to 128K, and gives vLLM examples with --mamba_ssm_cache_dtype float32." }, { "label": "NVIDIA Nemotron Nano 9B v2 served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/config.json", "source_type": "config", "supports": [ "model_family", "layers", "layer_pattern", "kv_heads", "head_dim", "mamba_recurrent_state", "max_context_tokens", "serving" ], "notes": "The served config records NemotronHForCausalLM, model_type nemotron_h, bfloat16, 56 layers, hybrid_override_pattern M-M-M-MM-M-M-M*-M-M-M*-M-M-M-M*-M-M-M-M*-M-MM-M-M-M-M-M-, 27 Mamba layers, 25 MLP-only layers, 4 attention layers, hidden_size 4480, intermediate_size 15680, 40 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 131072, sliding_window null, tie_word_embeddings false, vocab_size 131072, mamba_num_heads 128, mamba_head_dim 80, n_groups 8, conv_kernel 4, and ssm_state_size 128." }, { "label": "NVIDIA Nemotron Nano 12B v2 parent config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2/raw/f428df0ec725fed457b89cfca54dc26500fb88c1/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The parent config records the same NemotronH model type but different architecture geometry: 62 layers, hidden_size 5120, intermediate_size 20480, a different hybrid pattern, and a different served parameter total. This confirms the 9B v2 repo must be profiled from its own served config." }, { "label": "NVIDIA Nemotron H custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/modeling_nemotron_h.py", "source_type": "manual_review", "supports": [ "kv_adapter", "mamba_recurrent_state", "attention_kv_geometry" ], "notes": "Manual review found configuration_nemotron_h.py mapping M to Mamba2, * to attention, and - to MLP. HybridMambaAttentionDynamicCache stores attention key/value tensors with sequence length and Mamba conv_states plus ssm_states as fixed state. NemotronHMamba2Mixer derives conv_dim from Mamba inner width plus two grouped SSM projections, and the single-token update path runs causal_conv1d_update over that full hidden+B+C projection." }, { "label": "NVIDIA Nemotron Nano 9B v2 safetensors headers", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/resolve/6533e8de2c68e4536bf7c411d7a3ce5734111476/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "mamba_recurrent_state" ], "notes": "All four safetensors shard headers were range-read directly. Stored tensor bytes sum to 17.776454656 GB across 341 BF16 tensors, matching 8888227328 parameters and the index total_size. backbone.embeddings.weight has shape [131072, 4480] and contributes 587202560 parameters / 1.174405120 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. backbone.layers plus backbone.norm_f plus lm_head.weight total 8301024768 parameters / 16.602049536 GB. Layer grouping from the hybrid pattern gives Mamba tensors 7.962333696 GB, MLP tensors 7.024864000 GB, attention tensors 0.440437760 GB, final norm 0.000008960 GB, and lm_head 1.174405120 GB. Mamba conv1d.weight tensors have shape [12288, 1, 4], confirming the convolution-state width used in this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, 12B parent config comparison, custom runtime code, safetensors index, and direct range-read safetensors header byte grouping." }, "notes": "This self-contained profile deliberately uses current HF safetensors/header evidence instead of the older scraped catalog estimate, which overstated this repo as 12.31B active dense parameters and missed the fixed F32 SSM state serving requirement." }, { "id": "nvidia--nvlm-d-72b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/NVLM-D-72B", "title": "NVIDIA NVLM-D 72B", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's loaded NVLM-D 72B multimodal checkpoint.", "model_family": "nvlm-d-qwen2-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen2-72B-Instruct + OpenGVLab/InternViT-6B-448px-V1-5", "relation": "derived_package", "source": "Hugging Face model card/API metadata, served NVLM_D config, model.safetensors index, and direct safetensors header grouping", "config_compatible": true, "notes": "The card says NVLM-D 72B is trained from a Qwen2-72B-Instruct text backbone and an InternViT-6B vision encoder. The served config records the Qwen2ForCausalLM language geometry used for ordinary text decode: 80 layers, hidden size 8192, intermediate size 29568, 64 attention heads, 8 KV heads, 32768 max positions, dynamic RoPE scaling, untied token/output embeddings, and BF16 compute dtype. Vision and projector weights are part of the loaded checkpoint but are not swept during ordinary text-only decode after prefill." }, "architecture": { "canonical_architecture_id": "nvlm-d-72b-qwen2", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 79.379593344, "swept_params_b": 71.46049536, "auxiliary_resident_params_b": 7.919097984, "resident_weight_gb": 176.9021222, "swept_weight_gb": 156.34440192, "auxiliary_resident_weight_gb": 20.55772028, "resident_parameter_scope": "all linked safetensors payloads plus shard header/container bytes for the loaded NVLM-D checkpoint", "swept_parameter_scope": "ordinary text decode charges language_model.model.layers.*, language_model.model.norm.*, and language_model.lm_head.* tensors", "auxiliary_scope": "language_model.model.embed_tokens.*, vision_model.*, mlp1.*, and safetensors header/container bytes are resident but not swept for ordinary text-only decode", "notes": "Direct safetensors header grouping across all 46 shards found 176.901927168 GB of tensor payload and 0.000195032 GB of shard header/container overhead. Tensor payload groups are language layers 153.852968960 GB, language lm_head 2.491416576 GB, language norm 0.000016384 GB, language input embedding 2.491416576 GB, vision_model 16.824674560 GB, and mlp1 projector 1.241434112 GB." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served Qwen2 language config records 80 decoder layers, 8 KV heads, 128-dimensional key/value heads, and 32768 max positions. Bounds Engine v1 charges ordinary full-context BF16/FP16-sized K/V cache streams for text decode." }, "notes": "This profile is for ordinary text decode on a loaded multimodal checkpoint. It does not model image preprocessing, InternViT execution, projector prefill traffic, dynamic image tiling, or multimodal prefix KV policy." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.2285591894301557, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-nvlm-d-qwen2-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges stored BF16/F32 weights and BF16-sized KV traffic; custom-code kernels, activation traffic, image encoder execution, projector execution, and multimodal prefill are outside this ordinary text-decode bound.", "notes": "The checkpoint stores 70.308223104B BF16 parameters and 9.071370240B F32 parameters. Resident and swept byte fields are authoritative because the package mixes BF16 language tensors with F32/BF16 vision and projector tensors." }, "evidence": [ { "label": "NVIDIA NVLM-D 72B API metadata", "url": "https://huggingface.co/api/models/nvidia/NVLM-D-72B", "source_type": "model_card", "supports": [ "repo", "downloads", "license", "pipeline", "safetensors_parameter_split", "commit_sha" ], "notes": "At commit 11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f, the API reports a public non-gated Transformers image-text-to-text repo with custom_code, cc-by-nc-4.0 license, region:us, 159145 downloads, and safetensors parameters F32 9071370240, BF16 70308223104, total 79379593344." }, { "label": "NVIDIA NVLM-D 72B model card", "url": "https://huggingface.co/nvidia/NVLM-D-72B/raw/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "modality", "runtime_format" ], "notes": "The card identifies NVLM-D 72B as a decoder-only multimodal model with a Qwen2-72B-Instruct text-only LLM backbone and an InternViT-6B vision encoder. It says the model performs both vision-language and text-only tasks, uses trust_remote_code AutoModel loading, and is for non-commercial use." }, { "label": "NVIDIA NVLM-D 72B served config", "url": "https://huggingface.co/nvidia/NVLM-D-72B/raw/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "multimodal_components" ], "notes": "The served config records NVLM_D custom code with a Qwen2ForCausalLM language model: hidden_size 8192, intermediate_size 29568, 80 layers, 64 attention heads, 8 KV heads, max_position_embeddings 32768, rope_theta 1000000, dynamic RoPE scaling factor 3, tie_word_embeddings false, vocab_size 152064, and torch_dtype bfloat16. It also records an InternVisionModel vision config with 45 layers, hidden size 3200, image_size 448, patch_size 14, and dynamic image sizing." }, { "label": "NVIDIA NVLM-D 72B safetensors index", "url": "https://huggingface.co/nvidia/NVLM-D-72B/raw/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "shard_count", "tensor_grouping" ], "notes": "The index records total_size 176901927168 bytes across 46 safetensors shards and maps language_model, vision_model, and mlp1 projector tensors into the checkpoint." }, { "label": "NVIDIA NVLM-D 72B safetensors header range reads", "url": "https://huggingface.co/nvidia/NVLM-D-72B/tree/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "tensor_grouping" ], "notes": "Direct range reads of all 46 safetensors shard headers found 1556 tensors totaling 176.901927168 GB, with linked shard objects totaling 176.902122200 GB. Stored payload splits into BF16 140.616446208 GB / 70.308223104B parameters and F32 36.285480960 GB / 9.071370240B parameters. Ordinary text-decode swept language tensors total 156.344401920 GB across language layers, final norm, and lm_head. Resident-only tensors and header overhead total 20.557720280 GB: language input embedding 2.491416576 GB, vision_model 16.824674560 GB, mlp1 projector 1.241434112 GB, and shard header/container overhead 0.000195032 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served config, safetensors index, linked-object HEAD checks implied by direct header reads, and direct safetensors header grouping across all 46 shards." }, "notes": "Use this profile only for ordinary text decode on the loaded NVLM-D checkpoint. Multimodal prompt prefill and image-dependent prefix traffic require a separate multimodal workload adapter." }, { "id": "nvidia--phi-4-reasoning-plus-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Phi-4-reasoning-plus-NVFP4", "title": "NVIDIA Phi-4 Reasoning Plus NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 Phi-4 reasoning plus artifact.", "model_family": "phi3-dense", "base_model_proof": { "base_model": "microsoft/Phi-4-reasoning-plus", "relation": "quantized", "source": "Hugging Face card metadata, served config comparison, ModelOpt quantization config, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The NVIDIA card metadata identifies microsoft/Phi-4-reasoning-plus as the base model. Manual comparison found matching Phi3ForCausalLM geometry, context fields, RoPE settings, dtype, vocabulary, and untied embedding layout between the Microsoft BF16 base config and this NVIDIA ModelOpt artifact; the NVIDIA artifact adds ModelOpt quantization_config." }, "architecture": { "canonical_architecture_id": "phi-4-reasoning-plus", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.8437632, "swept_params_b": 7.32996096, "auxiliary_resident_params_b": 0.51380224, "resident_weight_gb": 9.723752, "swept_weight_gb": 8.69614752, "auxiliary_resident_weight_gb": 1.02760448, "resident_parameter_scope": "safetensors_header_stored_modelopt_nvfp4_u8_f8_e4m3_bf16_f32", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes transformer layers, model.norm.weight, and lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record a mixed ModelOpt representation totaling 9.723752 GB: U8 NVFP4 payload tensors, F8_E4M3 scale tensors, BF16 unquantized tensors, and tiny F32 scale tensors. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 10, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 40 layers, 10 KV heads, hidden_size 5120, 40 attention heads, and sliding_window null. The ModelOpt quantization config declares FP8 KV cache storage, so this full-context component is charged with one byte per scalar." }, "notes": "Dense Phi3ForCausalLM profile using the served NVIDIA ModelOpt config rather than deriving structure from the model name." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "tensorrt-llm-modelopt-nvfp4-fp8-kv-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored tensor bytes, including ModelOpt NVFP4 U8 payload tensors, F8 scale tensors, BF16/F32 side tensors, and FP8 KV bytes. Dequantization, activation traffic, TensorRT-LLM scheduling, and compute overhead are outside Bounds Engine v1.", "notes": "The model card targets TensorRT-LLM on NVIDIA Blackwell. config.json records quant_method modelopt and quant_algo NVFP4 with a static 8-bit float KV cache scheme; hf_quant_config.json independently records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Phi-4 reasoning plus NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Phi-4-reasoning-plus-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "serving", "weight_format", "total_params_b" ], "notes": "At commit cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b, the API reports a public MIT Model Optimizer repo with ModelOpt, FP4, modelopt, base_model:microsoft/Phi-4-reasoning-plus, and region:us tags. Current downloads are 472305. The API safetensors block records BF16 1028019200, F8_E4M3 851968000, U8 6815744000, and total 7843763200 storage-accounting tensor elements." }, { "label": "NVIDIA Phi-4 reasoning plus NVFP4 model card", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "hardware", "weight_format" ], "notes": "The card identifies the repo as the quantized version of Microsoft's Phi-4-reasoning-plus, says transformer-block linear operators are quantized to FP4 with TensorRT Model Optimizer, and lists TensorRT-LLM on NVIDIA Blackwell as the supported runtime target." }, { "label": "NVIDIA Phi-4 reasoning plus NVFP4 config", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Phi3ForCausalLM, bfloat16, hidden_size 5120, intermediate_size 17920, 40 layers, 40 attention heads, 10 KV heads, max_position_embeddings 32768, original_max_position_embeddings 32768, sliding_window null, tie_word_embeddings false, rope_theta 500000, vocab size 100352, and ModelOpt NVFP4 quantization with group_size 16, lm_head ignored, and static 8-bit float KV cache." }, { "label": "NVIDIA Phi-4 reasoning plus NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt 0.37.0.dev5, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and lm_head exclusion." }, { "label": "Microsoft Phi-4 reasoning plus base config", "url": "https://huggingface.co/microsoft/Phi-4-reasoning-plus/raw/69baf8528e1bcf05f475034d9e5dd32875ed125f/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible" ], "notes": "Manual comparison found no checked architecture differences between the Microsoft BF16 base config and this NVIDIA ModelOpt artifact after excluding quantization_config and transformers_version." }, { "label": "NVIDIA Phi-4 reasoning plus NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "dtype_split" ], "notes": "The index records total_parameters 7843763200 and total_size 9723752000 bytes across two shards. Range-read safetensors headers found 803 tensors totaling 9.723752 GB: U8 6.815744000 GB, BF16 2.056038400 GB, F8_E4M3 0.851968000 GB, and F32 0.000001600 GB. model.embed_tokens.weight has shape [100352, 5120] and contributes 0.513802240B parameters / 1.027604480 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 8.696147520 GB. Linked-object HEAD checks resolved both shards to 9.723839816 GB, leaving 87816 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the NVIDIA model card, served config, hf_quant_config, Microsoft base config comparison, safetensors index, direct safetensors shard header range reads, and local scrape row." }, "notes": "This profile supersedes the catalog estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged BF16 KV instead of the artifact's declared FP8 KV." }, { "id": "nvidia--qwen3-32b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Qwen3-32B-NVFP4", "title": "NVIDIA Qwen3 32B NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 package of Qwen3 32B.", "model_family": "qwen3-dense-nvfp4", "base_model_proof": { "base_model": "Qwen/Qwen3-32B", "relation": "quantized", "source": "Hugging Face model metadata, NVIDIA model card, public served config, ModelOpt quantization config, base config comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The API and model card identify Qwen/Qwen3-32B as the base model. Manual comparison found matching architecture, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, max-position embeddings, vocabulary size, tied-embedding setting, and full-attention settings between the current base config and this NVIDIA ModelOpt artifact. The NVIDIA artifact adds NVFP4/FP8 ModelOpt quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-32b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 17.159312384, "swept_params_b": 16.381400064, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 20.666169344, "swept_weight_gb": 19.110344704, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "safetensors_index_total_parameters_bf16_f8_e4m3_u8_storage_accounting", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, lm_head.weight output projection, and ModelOpt scale tensors", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The config records tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors. The API/index total_parameters count BF16, F8_E4M3, and U8 storage-accounting tensor elements; direct safetensors data_offsets drive byte traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The public config records 64 layers, 8 KV heads, 128 head dimension, 40960 max position embeddings, sliding_window null, and use_sliding_window false. config.json records a static 8-bit float KV cache scheme, and hf_quant_config.json records kv_cache_quant_algo FP8, so this profile charges one byte per KV scalar." }, "notes": "Dense Qwen3ForCausalLM profile using the served NVIDIA ModelOpt NVFP4 config and exact stored safetensors bytes." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "tensorrt-llm-modelopt-nvfp4-fp8-kv-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored tensor bytes, including ModelOpt NVFP4 U8 payload tensors, F8_E4M3 scale tensors, BF16 embeddings/lm_head/norms, tiny F32 scale tensors, and FP8 KV bytes. Dequantization, activation traffic, TensorRT-LLM scheduling, and compute overhead are outside Bounds Engine v1.", "notes": "config.json records quant_method modelopt, quant_algo NVFP4, group_size 16, lm_head ignored, and a static 8-bit float KV cache scheme. hf_quant_config.json independently records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Qwen3 32B NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-32B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "weight_format", "total_params_b", "commit_sha" ], "notes": "At commit 16426c6eb87be9e27c14cc9fb318f9c7a5f8588c, the API records a public non-gated Model Optimizer repo with safetensors, qwen3, nvidia, ModelOpt, quantized, FP4/fp4, base_model:Qwen/Qwen3-32B, Apache-2.0 license, 8-bit, modelopt, region:us, and 246588 downloads. The API safetensors block records BF16 1556501504, F8_E4M3 1950351360, U8 15602810880, and total 17159312384 storage-accounting tensor elements." }, { "label": "NVIDIA Qwen3 32B NVFP4 model card", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "hardware", "max_context_tokens" ], "notes": "The card describes the artifact as an FP4 quantized language model of Qwen3-32B, states context length up to 131K, lists TensorRT-LLM support, targets NVIDIA Blackwell, and says only linear operators within transformer blocks are quantized." }, { "label": "NVIDIA Qwen3 32B NVFP4 config", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen3ForCausalLM, bfloat16 runtime, hidden size 5120, intermediate size 25600, 64 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, vocab size 151936, tie_word_embeddings false, use_cache true, and ModelOpt NVFP4 quantization with group_size 16, lm_head ignored, and static 8-bit float KV cache." }, { "label": "NVIDIA Qwen3 32B NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt 0.35.0, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and lm_head exclusion." }, { "label": "Qwen3 32B BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison against the current Qwen/Qwen3-32B base config found matching checked architecture fields: Qwen3ForCausalLM, hidden_size 5120, intermediate_size 25600, 64 layers, 64 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 1000000, attention_bias false, and attention_dropout 0." }, { "label": "NVIDIA Qwen3 32B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "dtype_split" ], "notes": "The index records total_parameters 17159312384 and total_size 20666169344 bytes across five shards. Range-read safetensors headers found 2179 tensors totaling exactly 20.666169344 GB: U8 15.602810880 GB, BF16 3.113003008 GB, F8_E4M3 1.950351360 GB, and F32 0.000004096 GB. model.embed_tokens.weight has shape [151936, 5120] and contributes 0.777912320B parameters / 1.555824640 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 19.110344704 GB. Linked-object HEAD checks resolved the five shards to 20.666404224 GB, leaving 234880 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned NVIDIA model card, served ModelOpt NVFP4 config, hf_quant_config.json, generation config, current Qwen3 32B base config comparison, safetensors index, direct safetensors shard header range reads, and linked-object HEAD checks." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged BF16 KV instead of the artifact's declared FP8 KV." }, { "id": "nvidia--qwen3-5-397b-a17b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Qwen3.5-397B-A17B-NVFP4", "title": "NVIDIA Qwen3.5 397B A17B NVFP4", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's ModelOpt NVFP4 Qwen3.5 397B A17B serving artifact.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-397B-A17B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, hf_quant_config, and direct base-config comparison", "config_compatible": true, "notes": "The NVIDIA repo records Qwen/Qwen3.5-397B-A17B as its quantized base model. Manual comparison found matching audited text geometry between the pinned NVIDIA config and the pinned BF16 base config: Qwen3_5MoeForConditionalGeneration, 60 text layers, 15 full-attention layers, 45 DeltaNet linear-attention layers, 32 attention heads, 2 KV heads, 256 full-attention head dimension, 512 experts, 10 routed experts per token, one shared expert, 262144 max positions, and the same DeltaNet state geometry. The quantized repo adds ModelOpt NVFP4 and FP8-KV metadata; the only audited vision-geometry difference is a dtype label." }, "architecture": { "canonical_architecture_id": "qwen3-5-397b-a17b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 251.135201368, "main_resident_weight_gb": 234.997806712, "auxiliary_resident_weight_gb": 16.137394656, "fixed_weight_gb": 4.679448184, "routed_expert_weight_gb": 0.449840544, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embeddings, vision tower, and top-level MTP tensors", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024 and the model card states 10 routed plus 1 shared activated experts. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and tiny F32 scale tensors. Routed experts are byte-uniform across 512 expert indexes; routed_expert_weight_gb is the grouped ordinary-language routed tensor byte count divided by 512." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 15, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 60 layers with every fourth layer using full attention, giving 15 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.19316736, "read_gb_per_output_token": 0.19316736, "state_formula": "45 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and top-level MTP tensors in the package. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-sglang-modelopt-nvfp4-fp8-kv-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored NVFP4/FP8/BF16/F32 safetensors bytes and FP8 KV bytes. NVFP4 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The served config records a ModelOpt quantization config with an 8-bit float KV cache scheme, and hf_quant_config records quant_algo NVFP4 with kv_cache_quant_algo FP8. weight_bytes_per_param records the nominal NVFP4 weight payload; the audited adapter uses exact stored tensor bytes for resident and per-token weight traffic." }, "evidence": [ { "label": "NVIDIA Qwen3.5 397B A17B NVFP4 API metadata and model card", "url": "https://huggingface.co/api/models/nvidia/Qwen3.5-397B-A17B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit 12b28061efd8145b11226e4fab2528ee4f39ef09, the API records a public Apache-2.0 text-generation Model Optimizer artifact derived from Qwen/Qwen3.5-397B-A17B, with current downloads 488162, region:us, and no API safetensors parameter block. The card states the artifact is a ModelOpt NVFP4 quantization of Qwen3.5-397B-A17B, with 397B total parameters, 17B activated parameters, context length up to 262K, and SGLang/vLLM serving support." }, { "label": "NVIDIA Qwen3.5 397B A17B NVFP4 config", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/raw/12b28061efd8145b11226e4fab2528ee4f39ef09/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "kv_store_format", "kv_read_format", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 60 text layers, layer_types with every fourth layer full_attention, 15 full-attention layers, 45 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 512 experts, 10 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, top-level MTP tensors, and a ModelOpt quantization_config with kv_cache_scheme {dynamic:false, num_bits:8, type:'float'}." }, { "label": "NVIDIA Qwen3.5 397B A17B NVFP4 hf_quant_config", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/raw/12b28061efd8145b11226e4fab2528ee4f39ef09/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "ignored_quantized_modules" ], "notes": "The ModelOpt sidecar records quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes lm_head, shared experts, linear-attention modules, self-attention modules on full-attention layers, model.visual, and mtp.layers.0 from NVFP4 weight quantization." }, { "label": "Qwen3.5 397B A17B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/8472618112abcbd45acbcdc58436aff4233c23f7/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited top-level and text geometry fields between the NVIDIA NVFP4 config and BF16 base config. The base API records Apache-2.0 metadata, current downloads 420092, region:us, and safetensors parameters BF16 403397920304 plus F32 8640." }, { "label": "NVIDIA Qwen3.5 397B A17B NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/resolve/12b28061efd8145b11226e4fab2528ee4f39ef09/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 11 indexed shards. Stored tensors sum to the index total_size, 251.135201368 GB, across 371474 tensors: 193.27352832 GB U8, 24.15919104 GB F8_E4M3, 33.701744608 GB BF16, and 0.0007374 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 234.997806712 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 16.137394656 GB. Top-level MTP tensors account for 13.191136256 GB, visual tensors 0.91202096 GB, and the input embedding 2.03423744 GB. Ordinary-language routed expert tensors sum to 230.318358528 GB and divide exactly into 512 uniform expert indexes of 0.449840544 GB. Fixed ordinary text traffic sums to 4.679448184 GB. All index weight_map entries were found in the shard headers." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, served config, hf_quant_config, base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package does not fit on 128GB local hardware once runtime overhead is included." }, { "id": "nvidia--qwen3-6-35b-a3b-nvfp4", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Qwen3.6-35B-A3B-NVFP4", "title": "NVIDIA Qwen3.6 35B A3B NVFP4", "summary": "Audited memory-side bounds profile for the NVFP4 Qwen3.6 35B A3B serving artifact.", "model_family": "qwen3.6-moe", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and NVIDIA model card", "config_compatible": true, "notes": "The profile embeds the Qwen3.6 architecture and the NVIDIA NVFP4 serving representation." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 23.407580856, "main_resident_weight_gb": 19.808038104, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 1.688399064, "routed_expert_weight_gb": 0.07077984, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The model card states 8 routed plus 1 shared expert. The v1 adapter folds always-on shared expert traffic into the fixed active-parameter term.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The package mixes packed U8 NVFP4 tensors, F8 attention tensors, BF16 auxiliary tensors, and F32 scalar scales; routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The language stack repeats 3 linear-attention layers followed by 1 gated-attention layer across 40 layers, giving 10 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The NVFP4 artifact preserves the base Qwen3.6 text architecture, so quantizing weights and using FP8 KV for full-attention layers does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "This profile follows the Qwen row from the original bounds note by using NVFP4 weights and FP8 KV cache for the full-attention layers, but replaces the note's rounded weight bytes with exact safetensors-header byte accounting and includes the same audited DeltaNet fixed-state charge used by the base Qwen3.6 profile." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4", "dequantization_notes": "The memory-side bound charges stored NVFP4 weight bytes and FP8 KV bytes; compute and dequantization overheads are outside Bounds Engine v1.", "notes": "The NVIDIA card recommends FP8 KV cache for DGX Spark serving. weight_bytes_per_param records the nominal NVFP4 weight payload; the audited adapter uses exact stored tensor bytes for resident and per-token weight traffic." }, "evidence": [ { "label": "Qwen3.6 35B A3B model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "The overview states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 256 experts, and 8 routed plus 1 shared expert." }, { "label": "Qwen3.6 35B A3B config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, and 262144 max position embeddings." }, { "label": "NVIDIA Qwen3.6 NVFP4 model card", "url": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "weight_format", "kv_store_format", "kv_read_format" ], "notes": "The card identifies the artifact as a Model Optimizer NVFP4 quantization of Qwen3.6 35B A3B and recommends fp8 KV cache in its DGX Spark serving command." }, { "label": "NVIDIA Qwen3.6 NVFP4 safetensors headers", "url": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4/resolve/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The Hugging Face API reports repo SHA 491c2f1ea524c639598bf8fa787a93fed5a6fbce and safetensors dtype counts BF16 1825916784, F8_E4M3 3332177920, and U8 16423321600. Safetensors headers were range-read across the three shards. Stored tensor data sum to 23.407580856 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 19.808038104 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Main routed expert tensors sum to 18.11963904 GB, or 0.07077984 GB per expert index. Fixed ordinary text traffic sums to 1.688399064 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." }, { "label": "Original local frontier bounds note", "url": "https://github.com/osolmaz/onurclaw/blob/main/docs/2026-06-30-local-frontier-model-bounds.md", "source_type": "manual_review", "supports": [ "worked_example_parameters", "bounds_regression_target" ], "notes": "The original worked example supplied the Qwen3.6 NVFP4 architecture, FP8-KV serving setup, and rounded 35B/3B parameter-scale target. This audited profile keeps that serving setup but supersedes the rounded weight-byte assumption with safetensors-header bytes and separately charges the DeltaNet fixed state." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the model card, config, NVIDIA serving card, Hugging Face API metadata, range-read safetensors headers, original bounds note, and the Transformers qwen3_5 runtime implementation." }, "notes": "Production code must use this self-contained profile rather than inferring NVFP4 serving assumptions from the base Qwen repo name." }, { "id": "nvidia--qwen3-8b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Qwen3-8B-NVFP4", "title": "NVIDIA Qwen3 8B NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for NVIDIA's ModelOpt NVFP4 package of Qwen3 8B.", "model_family": "qwen3-dense-nvfp4", "base_model_proof": { "base_model": "Qwen/Qwen3-8B", "relation": "quantized", "source": "Hugging Face model metadata, NVIDIA model card, public served config, ModelOpt quantization config, current base config comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The API and model card identify Qwen/Qwen3-8B as the base model. Manual comparison found matching architecture, hidden size, intermediate size, layer count, attention heads, KV heads, head dimension, max-position embeddings, vocabulary size, tied-embedding setting, and full-attention settings between the current base config and this NVIDIA ModelOpt artifact. The NVIDIA artifact adds NVFP4/FP8 ModelOpt quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.717851648, "swept_params_b": 4.095521792, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 6.396932352, "swept_weight_gb": 5.15227264, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_index_total_parameters_bf16_f8_e4m3_u8_storage_accounting", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, lm_head.weight output projection, and ModelOpt scale tensors", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The config records tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors. The API/index total_parameters count BF16, F8_E4M3, and U8 storage-accounting tensor elements; direct safetensors data_offsets drive byte traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The public config records 36 layers, 8 KV heads, 128 head dimension, 40960 max position embeddings, sliding_window null, and use_sliding_window false. config.json records a static 8-bit float KV cache scheme, and hf_quant_config.json records kv_cache_quant_algo FP8, so this profile charges one byte per KV scalar." }, "notes": "Dense Qwen3ForCausalLM profile using the served NVIDIA ModelOpt NVFP4 config and exact stored safetensors bytes." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "tensorrt-llm-modelopt-nvfp4-fp8-kv-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored tensor bytes, including ModelOpt NVFP4 U8 payload tensors, F8_E4M3 scale tensors, BF16 embeddings/lm_head/norms, tiny F32 scale tensors, and FP8 KV bytes. Dequantization, activation traffic, TensorRT-LLM scheduling, and compute overhead are outside Bounds Engine v1.", "notes": "config.json records quant_method modelopt, quant_algo NVFP4, group_size 16, lm_head ignored, and a static 8-bit float KV cache scheme. hf_quant_config.json independently records quant_algo NVFP4 and kv_cache_quant_algo FP8." }, "evidence": [ { "label": "NVIDIA Qwen3 8B NVFP4 API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-8B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "weight_format", "total_params_b", "commit_sha" ], "notes": "At commit ccd10a893cbca613259517c3efe08e151ddf2b8e, the API records a public non-gated Model Optimizer repo with safetensors, qwen3, nvidia, ModelOpt, quantized, FP4/fp4, base_model:Qwen/Qwen3-8B, Apache-2.0 license, 8-bit, modelopt, region:us, and 150764 downloads. The API safetensors block records BF16 1244967936, F8_E4M3 434110464, U8 3472883712, and total 4717851648 storage-accounting tensor elements." }, { "label": "NVIDIA Qwen3 8B NVFP4 model card", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "hardware", "max_context_tokens" ], "notes": "The card describes the artifact as an FP4 quantized language model of Qwen3-8B, states context length up to 131K, lists TensorRT-LLM support, targets NVIDIA Blackwell, and says only linear operators within transformer blocks are quantized. The served config and this profile use max_position_embeddings 40960." }, { "label": "NVIDIA Qwen3 8B NVFP4 config", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen3ForCausalLM, bfloat16 runtime, hidden size 4096, intermediate size 12288, 36 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, vocab size 151936, tie_word_embeddings false, use_cache true, and ModelOpt NVFP4 quantization with group_size 16, lm_head ignored, and static 8-bit float KV cache." }, { "label": "NVIDIA Qwen3 8B NVFP4 quantization config", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "serving" ], "notes": "hf_quant_config records producer modelopt 0.35.0, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and lm_head exclusion." }, { "label": "Qwen3 8B BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/b968826d9c46dd6066d109eabc6255188de91218/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter" ], "notes": "Manual comparison against the current Qwen/Qwen3-8B base config found matching checked architecture fields: Qwen3ForCausalLM, hidden_size 4096, intermediate_size 12288, 36 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 1000000, attention_bias false, and attention_dropout 0." }, { "label": "NVIDIA Qwen3 8B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "dtype_split" ], "notes": "The index records total_parameters 4717851648 and total_size 6396932352 bytes across two shards. Range-read safetensors headers found 1227 tensors totaling exactly 6.396932352 GB: U8 3.472883712 GB, BF16 2.489935872 GB, F8_E4M3 0.434110464 GB, and F32 0.000002304 GB. model.embed_tokens.weight has shape [151936, 4096] and contributes 0.622329856B parameters / 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 5.152272640 GB. Linked-object HEAD checks resolved the two shards to 6.397066384 GB, leaving 134032 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned NVIDIA model card, served ModelOpt NVFP4 config, hf_quant_config.json, current Qwen3 8B base config comparison, safetensors index, direct safetensors shard header range reads, and linked-object HEAD checks." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as simple 0.5-byte NVFP4 weights and charged BF16 KV instead of the artifact's declared FP8 KV." }, { "id": "nvidia--qwen3-vl-235b-a22b-instruct-nvfp4-mlperf-inference-closed-v6-0", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0", "title": "NVIDIA Qwen3 VL 235B A22B Instruct NVFP4 MLPerf Inference Closed V6.0", "summary": "Audited memory-side text-decode bounds profile for NVIDIA's compressed-tensors NVFP4 Qwen3-VL 235B A22B Instruct MLPerf Inference Closed V6.0 package.", "model_family": "qwen3-vl-moe-nvfp4-mlperf", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-235B-A22B-Instruct", "relation": "quantized", "source": "Repository name, served compressed-tensors config comparison, quantization recipe, audited BF16 base profile, and safetensors header review", "config_compatible": true, "notes": "The repository has no model card metadata, but its name identifies Qwen3-VL-235B-A22B-Instruct and the pinned served config matches the audited BF16 base profile across checked text and vision geometry. The repo adds compressed-tensors NVFP4 quantization metadata while preserving the Qwen3-VL MoE architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-vl-235b-a22b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 135.254428064, "main_resident_weight_gb": 132.85699168, "auxiliary_resident_weight_gb": 2.397436384, "fixed_weight_gb": 5.114980288, "routed_expert_weight_gb": 0.997984464, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_nvfp4_u8_f8_bf16_f32", "traffic_scope": "ordinary text decode excludes resident visual tensors and input embeddings, and charges fixed language/logit tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config does not record a shared expert. Router/gate tensors are BF16 and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "The swept fixed subset includes non-expert model.language_model tensors except input embeddings, plus lm_head.weight. The compressed-tensors recipe ignores visual modules, lm_head, and all mlp.gate modules, so those tensors remain in the fixed or auxiliary BF16 buckets. Routed expert tensors and sidecars are byte-uniform across all 128 expert indexes and 94 layers." }, "kv_adapter": { "kind": "full_context", "layers": 94, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 94 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput or MLPerf harness overhead." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 1.0140298766631184, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "nvidia-compressed-tensors-nvfp4-qwen3-vl-moe-mlperf-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored compressed-tensors bytes: packed U8 payloads, F8_E4M3 tensors, BF16 ignored modules, and tiny F32 global scale tensors from safetensors headers. NVFP4 dequantization, activation traffic, compute overhead, MLPerf harness behavior, and vision prefill are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized Linear weights with group_size 16, local dynamic 4-bit float activations, ignored visual modules, ignored lm_head, ignored router/gate modules, BF16 runtime dtype, and kv_cache_scheme null. Exact resident and traffic byte fields drive production bounds." }, "evidence": [ { "label": "NVIDIA Qwen3 VL 235B A22B Instruct NVFP4 MLPerf API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0", "source_type": "model_card", "supports": [ "repo", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 4875dbf337c70bf4cf44b712375bab7bd77eab3e, the live API records a public non-gated repo with safetensors, qwen3_vl_moe, 8-bit, compressed-tensors, and region:us tags. The API has no model card data, license, library, or pipeline tag. Current downloads are 106913. The safetensors block reports BF16: 1871129328, F32: 72944, F8_E4M3: 14612430848, U8: 116899446784, total: 133383079904 storage-accounting elements." }, { "label": "NVIDIA Qwen3 VL 235B A22B Instruct NVFP4 MLPerf config", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/raw/4875dbf337c70bf4cf44b712375bab7bd77eab3e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "routed_experts", "routed_experts_per_token", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, qwen3_vl_moe, root tie_word_embeddings false, 94 text layers, hidden size 4096, intermediate size 12288, MoE intermediate size 1536, 64 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, decoder_sparse_step 1, norm_topk_prob true, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, BF16 text dtype, kv_cache_scheme null, and compressed-tensors nvfp4-pack-quantized Linear weights with 211 ignored module patterns." }, { "label": "NVIDIA Qwen3 VL 235B A22B Instruct NVFP4 MLPerf recipe", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/raw/4875dbf337c70bf4cf44b712375bab7bd77eab3e/recipe.yaml", "source_type": "manual_review", "supports": [ "serving", "weight_format", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The recipe records QuantizationModifier targets [Linear], scheme NVFP4, and ignore patterns for lm_head, visual/model.visual modules, and mlp.gate modules. The served config and safetensors headers remain authoritative for exact stored tensor bytes." }, { "label": "Qwen3 VL 235B A22B Instruct BF16 base profile and config comparison", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/710c13861be6c466e66de3f484069440b8f31389/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The already audited BF16 base profile records commit 710c13861be6c466e66de3f484069440b8f31389 with the same checked geometry used here: 94 text layers, hidden size 4096, 4 KV heads, 128 head dimension, 262144 context, 128 routed experts, 8 experts per token, untied embeddings, and a 27-layer visual tower. Manual comparison found the NVIDIA config preserving those architecture fields while adding compressed-tensors NVFP4 metadata." }, { "label": "Transformers 4.57.0 Qwen3 VL MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.0/src/transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "lm_head_layout", "embedding_layout" ], "notes": "Manual review for the BF16 base found Qwen3VLMoeAttention instantiates k_proj and v_proj using num_key_value_heads * head_dim and writes key_states/value_states through the cache path. Qwen3VLMoeTextSparseMoeBlock selects top_k experts from config.num_experts_per_tok. Qwen3VLMoeModel instantiates embed_tokens, and Qwen3VLMoeForConditionalGeneration instantiates a separate lm_head linear layer used for logits." }, { "label": "NVIDIA Qwen3 VL 235B A22B Instruct NVFP4 MLPerf safetensors index and shard headers", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/resolve/4875dbf337c70bf4cf44b712375bab7bd77eab3e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "serving", "embedding_layout" ], "notes": "Range-read headers across all 28 safetensors shards found 146712 tensors totaling 135.254428064 GB of tensor payload, matching the index total_size. Linked shard bytes total 135.274729272 GB, leaving 0.020301208 GB of safetensors header/container overhead outside tensor payloads. Payload bytes split into U8 116.899446784 GB, F8_E4M3 14.612430848 GB, BF16 3.742258656 GB, and F32 0.000291776 GB. model.language_model.embed_tokens.weight contributes 1.244659712 GB resident-only for ordinary decode. model.visual tensors total 1.152776672 GB resident-only. Ordinary main text/logit resident bytes total 132.856991680 GB. Fixed non-expert text/logit traffic totals 5.114980288 GB. Routed expert tensors plus sidecars total 127.742011392 GB, exactly 0.997984464 GB per routed expert across all 94 layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, served NVFP4 config, recipe.yaml, audited BF16 base profile/config comparison, official Transformers Qwen3 VL MoE implementation review, direct safetensors index resolution, and range-read safetensors shard headers across all 28 shards." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token fixed language/logit weights and expected distinct routed expert traffic, does not infer a model card/license/pipeline where the API omits them, and does not assume quantized KV cache without direct serving evidence." }, { "id": "ocicek--qwen3-6-27b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ocicek/Qwen3.6-27B-NVFP4", "title": "ocicek Qwen3.6 27B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the ocicek compressed-tensors NVFP4 package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, recipe, and served config comparison", "config_compatible": true, "notes": "The model card and API metadata identify this package as an NVFP4 quantized version of Qwen/Qwen3.6-27B. Manual comparison found matching audited text and vision geometry fields against the pinned Qwen/Qwen3.6-27B base config: Qwen3_5ForConditionalGeneration, language_model_only false, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, MTP draft head, and 262144 max position embeddings. The target adds compressed-tensors NVFP4 quantization metadata and a root dtype while preserving the base architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 20.558935392, "swept_weight_gb": 16.245279616, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "base logical Qwen3.6 parameters with exact stored NVFP4/BF16/F8/F32 safetensors bytes across LLM and MTP shards", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, the separate top-level MTP shard, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from both indexed safetensors files because the package mixes packed U8 NVFP4 weights, F8_E4M3 scales, F32 global scales, and unquantized BF16 tensors. Logical parameter counts follow the base Qwen3.6 model so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The target preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and NVFP4 dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and speculative MTP draft decoding are outside this ordinary text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes a resident vision tower plus a separate BF16 MTP shard. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored NVFP4 packed weights, F8 scales, F32 global scales, and unquantized BF16 tensors from safetensors headers. NVFP4 dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors format nvfp4-pack-quantized with 4-bit float weights and local dynamic 4-bit input activations, group_size 16, F8_E4M3 scale dtype, and kv_cache_scheme null. The card's vLLM quick-start uses --dtype auto and does not request FP8 KV cache, so attention KV is charged at BF16." }, "evidence": [ { "label": "ocicek Qwen3.6 27B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/ocicek/Qwen3.6-27B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 311613f3c18fe5dbc73a0fcaea7c721d258b1b4e, the API records a public non-gated vLLM repo with safetensors, qwen3_5, nvfp4, compressed-tensors, llm-compressor, text-generation, conversational, base_model:Qwen/Qwen3.6-27B, base_model:quantized:Qwen/Qwen3.6-27B, license:other, 8-bit, and region:us tags. Current downloads are 84773. The API safetensors block reports F32 992, BF16 3430871792, F8_E4M3 1521909760, U8 12175278080, total 17128060624 storage-accounting elements. The card identifies this as an NVFP4 quantized Qwen/Qwen3.6-27B package, says the vision tower and MTP draft head are preserved in BF16, records a 55.6 GB to 20.6 GB reduction, and gives a vLLM quick start for DGX Spark without FP8 KV." }, { "label": "ocicek Qwen3.6 27B NVFP4 config", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/raw/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The pinned config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors NVFP4 quantization, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config with depth 27 and hidden size 1152, quantization ignores for visual tensors, lm_head, and MTP linear modules, and kv_cache_scheme null." }, { "label": "ocicek Qwen3.6 27B NVFP4 recipe", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/raw/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope", "auxiliary_resident_scope" ], "notes": "The recipe applies NVFP4 to Linear targets and ignores lm_head, visual tensors, mlp.gate, and mlp.shared_expert_gate patterns. The resolved compressed-tensors config additionally lists the MTP linear modules in the ignore set, matching the model card statement that the MTP draft head is preserved in BF16." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences across the audited text_config and vision_config geometry fields between the target config and the pinned base config. The only checked root difference is the target's explicit root dtype field. Layer type arrays are identical." }, { "label": "ocicek Qwen3.6 27B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/resolve/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 20559284232 bytes across model-llm-nvfp4.safetensors and model-mtp-bf16.safetensors. Direct range-read headers found 2687 tensors totaling 20.558935392 GB of tensor payload, with linked file bytes totaling 20.559284232 GB and safetensors header/container overhead of 0.000348840 GB. Dtype bytes are U8 12.175278080 GB, BF16 6.861743584 GB, F8_E4M3 1.521909760 GB, and F32 0.000003968 GB. The ordinary text swept subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 16.245279616 GB. The resident-only subset totals 4.313655776 GB: model.language_model.embed_tokens.weight 2.542796800 GB, model.visual tensors 0.921460192 GB, and top-level MTP tensors 0.849398784 GB. lm_head.weight and model.language_model.embed_tokens.weight are separate BF16 tensors of shape 248320 x 5120." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, the model card, pinned NVFP4 config and recipe, pinned base config comparison, direct range-read safetensors headers for both indexed files, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted unquantized BF16 tensors, F8 scales, F32 scales, MTP, visual, embedding, and output-head storage. It is for ordinary text decode bounds after any multimodal prefill." }, { "id": "openai--gpt-oss-120b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "openai/gpt-oss-120b", "title": "OpenAI gpt-oss 120B MXFP4", "summary": "Audited memory-side bounds profile for the official MXFP4 gpt-oss-120b repo.", "model_family": "gpt-oss-moe", "architecture": { "canonical_architecture_id": "gpt-oss-120b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 65.248815744, "main_resident_weight_gb": 64.090548864, "auxiliary_resident_weight_gb": 1.15826688, "fixed_weight_gb": 3.096849024, "routed_expert_weight_gb": 0.47651328, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mxfp4_u8_bf16", "traffic_scope": "ordinary text decode through transformer layers and lm_head, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident, but ordinary decode traffic only looks up the selected token row rather than sweeping the full matrix", "notes": "Header-derived bytes are used because the package stores MXFP4 MoE block and scale tensors as U8 with BF16 attention, router, norm, bias, embedding, and lm_head tensors. Routed expert tensors are uniform by layer and expert index, so the expected-distinct adapter can use one per-expert byte total across all 36 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 18, "kv_heads": 8, "head_dim": 64, "notes": "Odd-numbered layers in the config use full_attention." }, { "kind": "sliding_window", "layers": 18, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "notes": "Even-numbered layers in the config use sliding_attention with a 128-token window." } ], "notes": "The config alternates sliding_attention and full_attention across 36 layers." }, "notes": "The model card describes gpt-oss-120b as a 117B-parameter MoE with 5.1B active parameters. The Hugging Face API reports 120.412337472 BF16/U8 safetensors parameters. The bounds profile uses exact stored-byte traffic for the served MXFP4 Hugging Face artifact rather than rounded marketed parameter counts." }, "serving": { "weight_format": "mxfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-mxfp4-moe-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored U8 MXFP4 block/scale bytes plus BF16 side tensors from safetensors headers. Dequantization and compute overhead are outside this memory-side bound.", "notes": "The repo API reports U8 and BF16 safetensors. The config marks self-attention, router, embeddings, and lm_head as modules not converted by MXFP4 quantization." }, "evidence": [ { "label": "OpenAI gpt-oss-120b model card and API metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-120b", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "serving", "marketed_total_params", "marketed_active_params" ], "notes": "The card records Apache-2.0 licensing, text-generation usage, 117B total parameters, 5.1B active parameters, and MXFP4 quantization of MoE weights. The API reports BF16 and U8 safetensors for the served repo, with SHA b5c939de8f754692c1647ca79fbf85e8c1e70f8a and lastModified 2025-08-26T17:25:03Z." }, { "label": "OpenAI gpt-oss-120b config", "url": "https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "max_context_tokens", "attention_pattern", "sliding_window", "kv_heads", "head_dim", "serving" ], "notes": "The config records GptOssForCausalLM, 36 hidden layers, 128 local experts, 4 experts per token, 131072 max position embeddings, alternating sliding/full attention, 128-token sliding windows, 8 KV heads, 64 head dimension, and MXFP4 quantization." }, { "label": "OpenAI gpt-oss-120b safetensors index and range-read headers", "url": "https://huggingface.co/openai/gpt-oss-120b/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across the 15 indexed shards. Stored tensors sum to 65.248815744 GB. Expert blocks, scales, and biases sum to 60.99369984 GB, or 0.47651328 GB per expert index across all 36 layers. The input embedding tensor is 1.15826688 GB resident-only for ordinary decode. The remaining transformer, router, norm, attention, bias, and lm_head tensors form 3.096849024 GB of fixed ordinary-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from model card/API metadata, config, safetensors index, and range-read safetensors headers." }, "notes": "This profile targets the official openai/gpt-oss-120b MXFP4 artifact. Quantized derivatives such as GGUF packages need their own self-contained profiles because their resident and traffic bytes differ." }, { "id": "openai--gpt-oss-20b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "openai/gpt-oss-20b", "title": "OpenAI gpt-oss 20B MXFP4", "summary": "Audited memory-side bounds profile for the official MXFP4 gpt-oss-20b repo.", "model_family": "gpt-oss-moe", "architecture": { "canonical_architecture_id": "gpt-oss-20b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.761264768, "main_resident_weight_gb": 12.602997888, "auxiliary_resident_weight_gb": 1.15826688, "fixed_weight_gb": 2.437381248, "routed_expert_weight_gb": 0.31767552, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mxfp4_u8_bf16", "traffic_scope": "ordinary text decode through transformer layers and lm_head, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident, but ordinary decode traffic only looks up the selected token row rather than sweeping the full matrix", "notes": "Header-derived bytes are used because the package stores MXFP4 MoE block and scale tensors as U8 with BF16 attention, router, norm, bias, embedding, and lm_head tensors. Routed expert tensors are uniform by layer and expert index, so the expected-distinct adapter can use one per-expert byte total across all 24 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 64, "notes": "Odd-numbered layers in the config use full_attention." }, { "kind": "sliding_window", "layers": 12, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "notes": "Even-numbered layers in the config use sliding_attention with a 128-token window." } ], "notes": "The config alternates sliding_attention and full_attention across 24 layers." }, "notes": "The model card describes gpt-oss-20b as a 21B-parameter MoE with 3.6B active parameters. The bounds profile uses exact stored-byte traffic for the served MXFP4 Hugging Face artifact rather than rounded marketed parameter counts." }, "serving": { "weight_format": "mxfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-mxfp4-moe-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored U8 MXFP4 block/scale bytes plus BF16 side tensors from safetensors headers. Dequantization and compute overhead are outside this memory-side bound.", "notes": "The repo API reports U8 and BF16 safetensors. The config marks self-attention, router, embeddings, and lm_head as modules not converted by MXFP4 quantization." }, "evidence": [ { "label": "OpenAI gpt-oss-20b model card and API metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-20b", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "serving", "marketed_total_params", "marketed_active_params" ], "notes": "The card records Apache-2.0 licensing, text-generation usage, 21B total parameters, 3.6B active parameters, and MXFP4 quantization of MoE weights. The API reports BF16 and U8 safetensors for the served repo." }, { "label": "OpenAI gpt-oss-20b config", "url": "https://huggingface.co/openai/gpt-oss-20b/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "max_context_tokens", "attention_pattern", "sliding_window", "kv_heads", "head_dim", "serving" ], "notes": "The config records GptOssForCausalLM, 24 hidden layers, 32 local experts, 4 experts per token, 131072 max position embeddings, alternating sliding/full attention, 128-token sliding windows, 8 KV heads, 64 head dimension, and MXFP4 quantization." }, { "label": "OpenAI gpt-oss-20b safetensors index and range-read headers", "url": "https://huggingface.co/openai/gpt-oss-20b/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across the three indexed shards. Stored tensors sum to 13.761264768 GB. Expert blocks, scales, and biases sum to 10.16561664 GB, or 0.31767552 GB per expert index across all 24 layers. The input embedding tensor is 1.15826688 GB resident-only for ordinary decode. The remaining transformer, router, norm, attention, bias, and lm_head tensors form 2.437381248 GB of fixed ordinary-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from model card/API metadata, config, safetensors index, and range-read safetensors headers." }, "notes": "This profile targets the official openai/gpt-oss-20b MXFP4 artifact. Quantized derivatives such as MLX or GGUF packages need their own self-contained profiles because their resident and traffic bytes differ." }, { "id": "openai--gpt-oss-safeguard-20b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "openai/gpt-oss-safeguard-20b", "title": "OpenAI gpt-oss-safeguard 20B MXFP4", "summary": "Audited memory-side text-decode bounds profile for the OpenAI gpt-oss-safeguard-20b MXFP4 safety-reasoning finetune.", "model_family": "gpt-oss-moe-safeguard", "base_model_proof": { "base_model": "openai/gpt-oss-20b", "relation": "finetune", "source": "Hugging Face model card base_model metadata, served config, OpenAI gpt-oss-20b base profile comparison, safetensors index metadata, linked shard HEAD checks, and direct safetensors header range reads", "config_compatible": true, "notes": "The model card/API metadata identify this as a finetune of openai/gpt-oss-20b. The served config preserves the base gpt-oss-20b memory-relevant architecture: 24 layers, 32 local experts, 4 experts per token, alternating sliding/full attention, 128-token sliding window, 8 KV heads, 64 key/value dimensions, and 131072 context length." }, "architecture": { "canonical_architecture_id": "gpt-oss-20b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.761264768, "main_resident_weight_gb": 12.602997888, "auxiliary_resident_weight_gb": 1.15826688, "fixed_weight_gb": 2.437381248, "routed_expert_weight_gb": 0.31767552, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_mxfp4_u8_bf16", "traffic_scope": "ordinary text decode through transformer layers and lm_head, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident, but ordinary decode traffic only looks up the selected token row rather than sweeping the full matrix", "notes": "Header-derived bytes are used because the package stores MXFP4 MoE block and scale tensors as U8 with BF16 attention, router, norm, bias, embedding, and lm_head tensors. Routed expert tensors are uniform by layer and expert index, so the expected-distinct adapter can use one per-expert byte total across all 24 layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "Odd-numbered layers in the config use full_attention." }, { "kind": "sliding_window", "layers": 12, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "kv_scalar_multiplier": 2, "notes": "Even-numbered layers in the config use sliding_attention with a 128-token window." } ], "notes": "The config alternates sliding_attention and full_attention across 24 layers, matching the official gpt-oss-20b architecture." }, "notes": "The model card describes gpt-oss-safeguard-20b as a 21B-parameter safety-reasoning model with 3.6B active parameters. The bounds profile uses exact stored-byte traffic for the served MXFP4 Hugging Face artifact rather than rounded marketed parameter counts." }, "serving": { "weight_format": "mxfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-mxfp4-moe-safeguard-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored U8 MXFP4 block/scale bytes plus BF16 side tensors from safetensors headers. Dequantization, safety-policy prompt overhead, tokenizer processing, and compute overhead are outside this memory-side bound.", "notes": "The repo API reports U8 and BF16 safetensors. The config marks self-attention, router, embeddings, and lm_head as modules not converted by MXFP4 quantization." }, "evidence": [ { "label": "OpenAI gpt-oss-safeguard-20b API metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-safeguard-20b", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "base_model_proof", "serving", "downloads", "logical_params" ], "notes": "The live HF API response at commit 8a11e17b25c973a24099d4016bf2e17dd7ec1574 records a public non-gated Apache-2.0 text-generation repo with base_model openai/gpt-oss-20b, base_model_relation finetune, 151082 downloads, region:us, endpoints_compatible, vLLM tag, and safetensors parameters BF16 1804459584, U8 19707494400, total 21511953984." }, { "label": "OpenAI gpt-oss-safeguard-20b model card", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "marketed_total_params", "marketed_active_params", "usage_scope" ], "notes": "The pinned card records Apache-2.0 licensing, text-generation pipeline, base_model openai/gpt-oss-20b, and finetune relation. It describes gpt-oss-safeguard-20b as a safety reasoning model built on gpt-oss, intended for safety use cases, with 21B parameters and 3.6B active parameters." }, { "label": "OpenAI gpt-oss-safeguard-20b config", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "max_context_tokens", "attention_pattern", "sliding_window", "kv_heads", "head_dim", "serving" ], "notes": "The config records GptOssForCausalLM, gpt_oss, 24 hidden layers, 32 local experts, 4 experts per token, 131072 max position embeddings, alternating sliding/full attention, 128-token sliding windows, 8 KV heads, 64 head dimension, hidden size 2880, untied embeddings, and MXFP4 quantization with self-attention, routers, embeddings, and lm_head excluded from conversion." }, { "label": "OpenAI gpt-oss 20B base profile", "url": "https://huggingface.co/openai/gpt-oss-20b/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "The existing audited openai/gpt-oss-20b profile records the same memory-relevant config geometry and exact stored-byte grouping. The safeguard finetune preserves the same tensor names, shapes, dtypes, and byte totals in the checked safetensors headers." }, { "label": "OpenAI gpt-oss-safeguard-20b safetensors index and range-read headers", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 13761264768 across model-00000-of-00002.safetensors, model-00001-of-00002.safetensors, and model-00002-of-00002.safetensors. Linked shard HEAD sizes are 4.792272488 GB, 4.798702184 GB, and 4.170342232 GB. Range-read headers across the three shards found 459 tensors totaling 13.761264768 GB: BF16 3.608919168 GB and U8 10.152345600 GB. Expert blocks, scales, and biases sum to 10.165616640 GB, or 0.317675520 GB per expert index across all 24 layers. The input embedding tensor is 1.158266880 GB resident-only for ordinary decode, lm_head.weight is 1.158266880 GB swept, and remaining non-embedding non-expert ordinary-decode traffic totals 2.437381248 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, the existing official gpt-oss-20b profile/config comparison, safetensors index metadata, linked shard HEAD checks, and direct range-read safetensors headers." }, "notes": "This profile targets the official openai/gpt-oss-safeguard-20b MXFP4 safety-reasoning finetune. It should not be reused for the base gpt-oss-20b repo or for GGUF/MLX derivatives because serving packages can change resident and traffic bytes." }, { "id": "opengvlab--internvl2-5-8b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "OpenGVLab/InternVL2_5-8B-AWQ", "title": "OpenGVLab InternVL2.5 8B AWQ", "summary": "Audited memory-side text-decode bounds profile for the AWQ 4-bit InternVL2.5 8B multimodal package.", "model_family": "internvl2.5-internlm2-dense-awq", "base_model_proof": { "base_model": "OpenGVLab/InternVL2_5-8B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served AWQ config, BF16 base config comparison, custom code hash comparison, PyTorch checkpoint index, and PyTorch zip data.pkl metadata", "config_compatible": true, "notes": "The AWQ repo records OpenGVLab/InternVL2_5-8B as its quantized base model. Manual comparison found matching InternVLChatModel, InternLM2ForCausalLM, and InternVisionModel geometry between the AWQ config and the BF16 base config; the AWQ artifact adds only the language-model AWQ quantization_config. The target and base custom configuration/modeling files have identical SHA-256 hashes." }, "architecture": { "canonical_architecture_id": "internvl2.5-8b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.075365376, "swept_params_b": 7.358685184, "auxiliary_resident_params_b": 0.716680192, "resident_weight_gb": 5.818062848, "swept_weight_gb": 4.384702464, "auxiliary_resident_weight_gb": 1.433360384, "resident_parameter_scope": "logical InternVL2.5 8B parameter count represented by the PyTorch AWQ package", "swept_parameter_scope": "ordinary text decode excludes language_model.model.tok_embeddings.weight input lookup, vision_model tensors, and mlp1 projector tensors while including language layers, language_model.model.norm.weight, and language_model.output.weight", "auxiliary_scope": "vision_model, mlp1 projector, and language_model.model.tok_embeddings.weight are resident for the multimodal package but not swept as full matrices for each ordinary text decode token", "notes": "Bounds use exact stored bytes from PyTorch zip data.pkl tensor metadata because this AWQ package ships three pytorch_model*.bin shards rather than safetensors. Stored tensors are HalfStorage and IntStorage. Logical parameter counts match the BF16 base model geometry: packed qweight tensors represent 4-bit logical matrices, while qzeros and scales are storage/runtime overhead." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The custom InternLM2 attention code caches standard K and V tensors for all 32 language layers and the served config has no sliding-window field. Head dimension is hidden_size 4096 divided by 32 attention heads." }, "notes": "InternVLChatModel is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder or projector throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7204705393629981, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "lmdeploy-awq-gemm-internvl2.5-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed I32 qweight/qzeros tensors, FP16 scales, and unquantized FP16 multimodal/text tensors from PyTorch checkpoint metadata. AWQ dequantization, activation traffic, vision prefill, projector prefill, kernel choice, and compute overhead are outside Bounds Engine v1.", "notes": "The top-level config records fp16 true and torch_dtype float16, while the nested InternLM2 config retains bfloat16/use_bfloat16 metadata. Both imply two bytes per KV scalar for this memory bound. The served language quantization_config records AWQ 4-bit GEMM, group_size 128, zero_point true." }, "evidence": [ { "label": "InternVL2.5 8B AWQ API metadata", "url": "https://huggingface.co/api/models/OpenGVLab/InternVL2_5-8B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "downloads", "pipeline", "commit_sha" ], "notes": "At commit 14e0a7932fa4d629e19209f7dd1d4003690e2b64, the API records a public non-gated MIT image-text-to-text Transformers repo with internvl_chat, internvl, custom_code, multilingual, base_model OpenGVLab/InternVL2_5-8B, base_model:quantized, and region:us tags. Current downloads were 304286 when audited. The repo ships PyTorch .bin shards, not safetensors." }, { "label": "InternVL2.5 8B AWQ model card", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ/raw/14e0a7932fa4d629e19209f7dd1d4003690e2b64/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "architecture", "serving" ], "notes": "The card identifies this as the AWQ package of InternVL2_5-8B, lists the vision part as InternViT-300M-448px-V2_5 and the language part as internlm2_5-7b-chat, describes the ViT-MLP-LLM architecture, and gives LMDeploy serving examples for OpenGVLab/InternVL2_5-8B-AWQ." }, { "label": "InternVL2.5 8B AWQ served config", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ/raw/14e0a7932fa4d629e19209f7dd1d4003690e2b64/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The config records InternVLChatModel, force_image_size 448, downsample_ratio 0.5, max_dynamic_patch 12, ps_version v2, top-level fp16 true and torch_dtype float16, InternLM2ForCausalLM language model with hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, 32768 max position embeddings, dynamic RoPE factor 2, tie_word_embeddings false, and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. The vision config records an InternVisionModel with hidden size 1024, 24 layers, 16 heads, 448 image size, and 14 patch size." }, { "label": "InternVL2.5 8B BF16 base config and custom code comparison", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B/raw/e9e4c0dc1db56bfab10458671519b7fa3dd29463/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison against the BF16 base found matching top-level InternVL fields, language geometry, vision geometry, dynamic RoPE, context length, and untied embedding setting. The base API records 8075365376 BF16 safetensors parameters. The target and base custom files have identical SHA-256 hashes: configuration_internlm2.py 9f9e14d64be7b81068e8421fd27e1c466b3fb6aeb67725ca328e56083c3f7a60, modeling_internlm2.py c014d933c6afdf7414791c1859c692ca4b49c925688bcc00fbd1658ef75fb0ba, configuration_internvl_chat.py 20bd8ddcf5c51f9fded0e3191db2bc00a7012fd3f7912dea4277fa24f17b6e32, modeling_internvl_chat.py 1915c32d984aed6750e911884ad59f92504f73b799c7443121c67b9cb21fe347, configuration_intern_vit.py 22dd687a5f142985407675fc6b1dc688c2ded93e54192543ecf28ef495a6efa8, and modeling_intern_vit.py 1978e330ea5533dda4d1379096aded271fc741347fc48cb4924846b56a2a6a4b." }, { "label": "InternVL2.5 8B AWQ PyTorch checkpoint index and zip metadata", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ/raw/14e0a7932fa4d629e19209f7dd1d4003690e2b64/pytorch_model.bin.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The PyTorch index maps 893 tensors across three shards and records metadata.total_size 5818062848 bytes. Range-reading the data.pkl metadata at the front of each PyTorch zip shard found exactly 5818062848 bytes of tensor storage: IntStorage 3516923904 bytes and HalfStorage 2301138944 bytes. Stored scope split is language_model 5.142896640 GB, vision_model 0.608024576 GB, and mlp1 projector 0.067141632 GB. language_model.model.tok_embeddings.weight is HalfStorage [92553, 4096] and contributes 0.758194176 GB resident-only. language_model.output.weight is stored separately with the same shape and remains in swept decode traffic. Language layer tensors plus final norm plus output head total 4.384702464 GB swept traffic. Resident-only vision, projector, and input embedding tensors total 1.433360384 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served AWQ config, BF16 base config comparison, custom code hash comparison, PyTorch checkpoint index, linked-object HEAD checks, and range-read PyTorch zip data.pkl tensor metadata." }, "notes": "This profile supersedes the generated ideal 4-bit estimate by using exact stored PyTorch checkpoint bytes and by excluding only resident multimodal/projector/input-embedding tensors from ordinary cached text-decode traffic." }, { "id": "palmfuture--qwen3-6-35b-a3b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4", "title": "palmfuture Qwen3.6 35B A3B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the palmfuture GPTQ Int4 package of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The GPTQ artifact records Qwen/Qwen3.6-35B-A3B as its quantized base model. Manual comparison found no differences in the checked top-level, text, vision, MoE, attention, context, and DeltaNet state geometry between this config and the BF16 base config. The target adds GPTQ quantization metadata and split MTP safetensors while preserving the base architecture." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.403152608, "main_resident_weight_gb": 20.803609856, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.06610944, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_gptq_i32_f16_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary non-speculative text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512 and the model card states top-8 routed experts. Shared expert tensors are kept BF16 by the GPTQ dynamic exclusion list and are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes packed GPTQ I32 tensors, F16 GPTQ scale tensors, and BF16 excluded tensors. The quantization dynamic list excludes attention, routers, shared experts, visual tensors, MTP tensors, embeddings, and lm_head. Routed expert tensors are byte-uniform across all 256 expert indexes after summing qweight, qzeros, g_idx, and scales tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers, following the palmfuture vLLM/SGLang serving commands, plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary non-speculative text decode after any multimodal prefill; speculative MTP decode requires a separate workload path." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6787737247403844, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-sglang-gptq-fp8-kv-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ/BF16/F16 safetensors bytes and FP8 full-attention KV bytes. GPTQ dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and recurrent-state writes are outside this memory-side bound.", "notes": "The config records bfloat16 model dtype and GPTQ 4-bit quantization. The model card production commands explicitly request FP8 KV cache for both SGLang and vLLM, so this profile charges full-attention KV as FP8 while keeping the DeltaNet fixed state charge from the audited Qwen3.6 implementation." }, "evidence": [ { "label": "palmfuture Qwen3.6 35B A3B GPTQ Int4 model card and API metadata", "url": "https://huggingface.co/api/models/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit d1fef185160f938fca00c3c664f21250dd544d63, the API records an Apache-2.0 image-text-to-text GPTQ artifact derived from Qwen/Qwen3.6-35B-A3B with qwen3_5_moe, gptq, int4, mtp, speculative-decoding, endpoints_compatible, and region:us tags. Current downloads were 449506 when audited, and API safetensors parameters were BF16 3739567984, I32 32212254720, total 35951822704. The card states GPTQ v2 4-bit quantization with group_size 128 and symmetric quantization, 35B total / 3B active MoE base, 256 experts with top-8 routing, 40 hidden layers, 262144 context length, 24.4 GB quantized size including MTP weights, BF16 MTP with 785 keys, and production SGLang/vLLM examples using FP8 KV cache." }, { "label": "palmfuture Qwen3.6 35B A3B GPTQ Int4 config", "url": "https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4/raw/d1fef185160f938fca00c3c664f21250dd544d63/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, bfloat16 root/text/vision dtype, qwen3_5_moe_text, 40 text layers, layer_types with every fourth layer full_attention, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, vocab size 248320, a 27-layer vision config, mtp_num_hidden_layers 1, and GPTQ quantization with bits 4, pack_dtype int32, group_size 128, desc_act false, symmetric quantization, and dynamic exclusions for attention, routers, MTP, shared experts, visual tensors, lm_head, and input embeddings." }, { "label": "Qwen3.6 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences across 30 checked architecture fields between this GPTQ config and the Qwen/Qwen3.6-35B-A3B base config at commit 995ad96eacd98c81ed38be0c5b274b04031597b0. The checked fields covered top-level architecture and tie settings, text layer pattern, hidden size, attention and linear-attention geometry, expert counts, context length, vocab size, and vision geometry. The BF16 base API reports safetensors parameters BF16 35951822704." }, { "label": "palmfuture Qwen3.6 35B A3B GPTQ Int4 safetensors headers", "url": "https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4/resolve/d1fef185160f938fca00c3c664f21250dd544d63/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across the six indexed model shards plus mtp.safetensors. Tensor payload spans sum to 24.403152608 GB across 124611 tensors: I32 16.420700160 GB, BF16 7.479135968 GB, and F16 0.503316480 GB. The index metadata total_size records 24.420021632 GB, which overstates the header payload sum by 0.016869024 GB; this profile uses the direct tensor payload spans. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 20.803609856 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB: visual 0.893142496 GB, MTP 1.689281536 GB, and input embedding 1.017118720 GB. Routed expert tensors sum to 16.924016640 GB, or 0.066109440 GB per expert index across all 256 experts. Fixed ordinary text traffic sums to 3.879593216 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, served GPTQ config, base config/API comparison, direct safetensors header byte grouping across indexed model shards and mtp.safetensors, and the existing Transformers qwen3_5 runtime implementation review used by sibling Qwen3.6 profiles." }, "notes": "This profile is for ordinary non-speculative text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "prithivmlmods--gemma-4-e4b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "prithivMLmods/gemma-4-E4B-it-NVFP4", "title": "prithivMLmods Gemma 4 E4B IT NVFP4", "summary": "Audited memory-side text-decode bounds profile for the prithivMLmods NVFP4 compressed-tensors package of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model metadata, served config comparison, quantization recipe, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo metadata records google/gemma-4-E4B-it as its quantized base model. Manual comparison found no differences across the checked text, vision, audio, context, tied-embedding, and attention geometry fields between the prithivMLmods config and the pinned Google base config, except that the prithivMLmods text_config explicitly adds moe_intermediate_size null. The artifact adds compressed-tensors NVFP4 quantization metadata while preserving the base architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.905065317, "swept_params_b": 2.937316005, "auxiliary_resident_params_b": 3.967749312, "resident_weight_gb": 11.544469952, "swept_weight_gb": 3.608971328, "auxiliary_resident_weight_gb": 7.935498624, "resident_parameter_scope": "safetensors_header_stored_nvfp4_u8_fp8_bf16_f32_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.language_model.embed_tokens.weight input lookup and model.language_model.embed_tokens_per_layer.weight, and includes language-model layer tensors plus lm_head.weight output projection", "auxiliary_scope": "model.language_model.embed_tokens.weight, model.language_model.embed_tokens_per_layer.weight, model.audio_tower, model.embed_audio, model.vision_tower, and model.embed_vision are resident for token lookup, multimodal prefill, and PLE packaging but not swept as full matrices for each ordinary text decode token", "notes": "Header-derived bytes are used because the artifact stores packed U8 NVFP4 language weights, F8_E4M3 scale tensors, tiny F32 global scales, BF16 embeddings, BF16 lm_head, and unquantized BF16 audio/vision modules. The config records tie_word_embeddings true, but the safetensors file stores separate BF16 model.language_model.embed_tokens.weight and lm_head.weight tensors; resident bytes include both stored tensors, while swept decode traffic charges one full BF16 vocabulary projection rather than double-counting both. The large per-layer embedding table remains resident-only for ordinary decode, matching the audited Google Gemma 4 E4B PLE convention." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, with separate K and V streams." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 35 layers use 512-token local sliding-window attention with separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The compressed-tensors config has kv_cache_scheme null, so this profile charges BF16 KV cache bytes like the Google base profile. Audio, image, and video prefill throughput is outside this ordinary text-decode profile." }, "notes": "Gemma4ForConditionalGeneration is multimodal and includes audio and vision towers. This profile models ordinary text decode after any multimodal prefill, not encoder throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 1.6718842504759466, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-gemma4-e4b-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed U8 NVFP4 tensors, F8_E4M3 scales, F32 scalar scales, BF16 embeddings, BF16 lm_head, BF16 norms, and unquantized multimodal tensors from the safetensors header. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors NVFP4 pack quantization with group size 16 and kv_cache_scheme null. The recipe ignores lm_head, vision, audio, embed_vision, and embed_audio modules, so KV remains BF16 and the ignored modules stay in their stored BF16 representation." }, "evidence": [ { "label": "prithivMLmods Gemma 4 E4B IT NVFP4 API metadata", "url": "https://huggingface.co/api/models/prithivMLmods/gemma-4-E4B-it-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA bd4ca527287ab7de27f32ec80fa6235d7c4b08e0, the API records a public Apache-2.0 any-to-any repo with base_model google/gemma-4-E4B-it, base_model:quantized metadata, compressed-tensors, vLLM, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 243514. The API safetensors block reports F32: 379, BF16: 4639402570, F8_E4M3: 251740160, U8: 2013921280, and total: 6905064389 storage-accounting elements." }, { "label": "prithivMLmods Gemma 4 E4B IT NVFP4 model card", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "quantized_module_scope" ], "notes": "The card describes this as an NVFP4-compressed evolution of google/gemma-4-E4B-it using F32, BF16, F8_E4M3, and U8 precision formats. It repeats the Gemma 4 E4B architecture summary: 4.5B effective parameters, 8B total with per-layer embeddings, 42 layers, 512-token sliding window, 128K context, and multimodal text/image/audio support." }, { "label": "prithivMLmods Gemma 4 E4B IT NVFP4 config", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/raw/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors NVFP4 pack quantization with 4-bit float weights and group_size 16, kv_cache_scheme null, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "prithivMLmods Gemma 4 E4B IT NVFP4 recipe", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/raw/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/recipe.yaml", "source_type": "config", "supports": [ "serving", "quantization", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier targets [Linear], scheme NVFP4A16, and ignores lm_head, vision_tower, audio_tower, embed_vision, and embed_audio modules." }, { "label": "Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no profile-relevant architecture differences between the prithivMLmods config and the pinned Google base config after excluding quantization metadata, chat template fields, and Transformers version. The only remaining checked diff is moe_intermediate_size null in the prithivMLmods text_config." }, { "label": "prithivMLmods Gemma 4 E4B IT NVFP4 safetensors header", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/resolve/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The linked object is 11.544852744 GB with a 382784-byte safetensors header. Range-reading the single header found 2889 tensors with payloads totaling 11.544469952 GB: BF16 9.278806996 GB, U8 2.013921280 GB, F8_E4M3 0.251740160 GB, and F32 0.000001516 GB. Header tensor elements total 6.905065317B. model.language_model tensors total 9.246115904 GB. model.language_model.embed_tokens.weight is BF16 [262144, 2560] and contributes 0.671088640B elements / 1.342177280 GB resident-only for ordinary decode. lm_head.weight is a separate BF16 tensor of the same shape and remains in swept decode traffic. model.language_model.embed_tokens_per_layer.weight is BF16 [262144, 10752] and contributes 2.818572288B elements / 5.637144576 GB resident-only for ordinary decode. Audio/embed_audio tensors total 0.617514496 GB, and vision/embed_vision tensors total 0.338662272 GB. Ordinary text swept traffic, defined as language_model tensors excluding embed_tokens and embed_tokens_per_layer plus lm_head, totals 2.937316005B elements / 3.608971328 GB. Auxiliary resident tensors, defined as input embedding plus per-layer embedding plus audio plus vision, total 3.967749312B elements / 7.935498624 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served compressed-tensors config, quantization recipe, pinned Google base config comparison, linked-object tree size, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the generated metadata estimate, which treated the artifact as an ideal 0.5-byte dense model and missed BF16 embeddings, the separate BF16 lm_head, F8 scale tensors, F32 scalar scales, multimodal towers, per-layer embedding residency, and hybrid Gemma 4 KV traffic." }, { "id": "pytorch--gemma-3-27b-it-awq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "pytorch/gemma-3-27b-it-AWQ-INT4", "title": "PyTorch Gemma 3 27B IT AWQ INT4", "summary": "Audited memory-side text-decode bounds profile for the PyTorch TorchAO AWQ INT4 package of Gemma 3 27B IT.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-27b-it", "relation": "quantized", "source": "Hugging Face API metadata, model card, served config, and PyTorch archive metadata", "config_compatible": false, "notes": "The repo metadata and model card identify google/gemma-3-27b-it as the quantized base. The base repo remains gated in this audit environment, so this profile uses the public PyTorch AWQ served config and archive metadata directly rather than copying the base config." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.43240664, "swept_params_b": 27.009346304, "auxiliary_resident_params_b": 0.423060336, "resident_weight_gb": 17.2737272, "swept_weight_gb": 16.427606528, "auxiliary_resident_weight_gb": 0.846120672, "resident_parameter_scope": "pytorch_zip64_pickle_logical_torchao_awq_int4_bf16", "swept_parameter_scope": "ordinary text decode charges TorchAO Int4Tensor language layer entries, language norm, and the tied embedding/output storage exposed as language_model.lm_head.weight", "auxiliary_scope": "vision_tower and multi_modal_projector tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "Byte-range ZIP64 central-directory reads plus data.pkl pickle metadata recover all 1248 state-dict entries without downloading the tensor payloads. Storage entries total 17.273727200 GB, matching pytorch_model.bin.index.json total_size. The language_model.model.embed_tokens.weight and language_model.lm_head.weight entries share one 2.819260416 GB BF16 storage; this profile charges that shared storage once in resident memory and once in swept output-projection traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 62 text layers with full_attention at indexes 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59." }, { "kind": "sliding_window", "layers": 52, "kv_heads": 16, "head_dim": 128, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 52 text layers use the config's 1024-token sliding_attention window." } ], "notes": "Layered KV models ordinary text decode after any image prefill. The config records no KV quantization scheme, so the profile charges BF16 K and V streams." }, "notes": "Gemma3ForConditionalGeneration is multimodal. This profile models ordinary generated text-token decode after any image prefill; vision encoder, image preprocessing, and prompt prefill throughput are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6296832584426869, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-torchao-awq-int4-pytorch-bin-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact TorchAO archive storage bytes from ZIP64 central-directory entries and data.pkl tensor metadata. TorchAO dequantization, activation traffic, H100-specific INT4 kernels, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The model card describes AWQ-INT4 quantization with TorchAO and vLLM. The saved artifact uses PyTorch .bin archives rather than safetensors; the profile audits the archive metadata directly." }, "evidence": [ { "label": "PyTorch Gemma 3 27B AWQ INT4 API metadata", "url": "https://huggingface.co/api/models/pytorch/gemma-3-27b-it-AWQ-INT4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "artifact_format" ], "notes": "At commit 0e6f8915f5b53333c13bbd48fc6ead75cc7625d7, the current API reports a public non-gated Apache-2.0 image-text-to-text repo with transformers, pytorch, torchao, gemma3, region:us, base_model google/gemma-3-27b-it, base_model:quantized:google/gemma-3-27b-it, 184232 downloads, no safetensors block, no GGUF block, and four PyTorch .bin shards." }, { "label": "PyTorch Gemma 3 27B AWQ INT4 config", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/raw/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "layer_pattern", "max_context_tokens", "serving", "vision_geometry" ], "notes": "The config records Gemma3ForConditionalGeneration, gemma3_text, 62 text layers, hidden size 5376, intermediate size 21504, 32 attention heads, 16 KV heads, 128 head dimension, 131072 max positions, 1024-token sliding window, full_attention every sixth layer, and a 27-layer SigLIP vision tower. The TorchAO quantization_config targets language-model MLP and self-attention projection modules with AWQ Int4WeightOnlyConfig group_size 128, TinyGEMM, plain packing, TensorCoreTiledLayout inner_k_tiles 8, and no zero point." }, { "label": "PyTorch Gemma 3 27B AWQ INT4 model card", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/raw/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "runtime_format", "artifact_format" ], "notes": "The card identifies the artifact as an AWQ-INT4 quantization of google/gemma-3-27b-it, recommends vLLM nightly plus TorchAO, says AWQ only works for H100 INT4 so far, and shows the push_to_hub recipe using safe_serialization=False." }, { "label": "PyTorch Gemma 3 27B AWQ INT4 PyTorch index", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/raw/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7/pytorch_model.bin.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "resident_weight_gb", "artifact_format" ], "notes": "The PyTorch index records total_parameters 27432406640, total_size 17273727200 bytes, 1248 state-dict entries, and four PyTorch .bin shards." }, { "label": "PyTorch Gemma 3 27B AWQ INT4 archive metadata", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/tree/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_scope", "weight_format" ], "notes": "Byte-range reads of all four ZIP64 central directories and uncompressed data.pkl entries found 17273727200 bytes of tensor storage and 1248 state-dict entries. Storage dtype split is CharStorage 12.799180800 GB and BFloat16Storage 7.293806816 GB. Grouped storage bytes are language MLP 11.427268608 GB, language self-attention 2.178400256 GB, other language layer tensors 0.002666496 GB, language norm 0.000010752 GB, tied embedding/output storage 2.819260416 GB, vision tower 0.833732064 GB, and multimodal projector 0.012388608 GB. The only shared storage is the tied embedding/output projection blob referenced by both language_model.model.embed_tokens.weight and language_model.lm_head.weight." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes the repeating local/global attention pattern as five local attention layers followed by one global attention layer, matching the served config's full_attention entries every sixth layer." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, PyTorch index, linked object sizes, direct ZIP64 central-directory range reads, direct data.pkl metadata parsing, and Gemma 3 local/global attention documentation." }, "notes": "This self-contained profile supersedes the generated ideal 0.5-byte-per-parameter estimate. It uses exact PyTorch archive storage bytes and Gemma 3 layered local/global KV for production bounds." }, { "id": "quanttrio--deepseek-v3-2-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/DeepSeek-V3.2-AWQ", "title": "QuantTrio DeepSeek V3.2 AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ package of DeepSeek V3.2.", "model_family": "deepseek-v32-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V3.2", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The AWQ artifact records deepseek-ai/DeepSeek-V3.2 as its quantized base model. Manual comparison found no differences in 25 checked architecture fields between the AWQ config and the official DeepSeek V3.2 config. The target changes quantization metadata from FP8 to AWQ while preserving the model geometry, bundled inference config, and bundled inference code." }, "architecture": { "canonical_architecture_id": "deepseek-v3-2", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 361.909457408, "main_resident_weight_gb": 352.008032512, "auxiliary_resident_weight_gb": 9.901424896, "fixed_weight_gb": 12.281991424, "routed_expert_weight_gb": 1.327054848, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_f16_f32", "traffic_scope": "ordinary decode through main layers 0-60, embeddings, norm, and lm_head, excluding resident-only next-token prediction layer 61", "auxiliary_scope": "model.layers.61 tensors are resident for the checkpoint but excluded from ordinary non-speculative decode traffic", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes packed AWQ I32 tensors, F16 side tensors, and small F32 tensors. Expected-distinct routing is applied to the 256 byte-uniform routed expert indexes across main MoE layers 3-60." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Conservative cache allocation coefficient from the bundled official inference config/code: 61 layers * ((kv_lora_rank 512 FP8 bytes + 4 F32 scale values) + qk_rope_head_dim 64 BF16 bytes + (index_head_dim 128 FP8 bytes + 1 F32 scale value)) * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.048068, "notes": "Bounds Engine v1 conservatively charges the same full-context MLA plus DSA index cache bytes per active read token. The V3.2 top-2048 DSA sparse read savings are not modeled by this linear formula." }, "notes": "DeepSeek V3.2 uses MLA plus DeepSeek Sparse Attention. The target AWQ repo bundles byte-identical official inference code/config, which keeps an MLA latent cache, a rope cache, and an FP8 indexer key cache." }, "notes": "The checkpoint has 61 main hidden layers and one next-token-prediction layer. Ordinary text decode excludes the auxiliary next-token-prediction layer from per-token traffic but keeps it resident." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5280610534069046, "kv_store_format": "mla_fp8_bf16_rope_dsa_fp8_index", "kv_store_bytes_per_scalar": 1, "kv_read_format": "conservative_full_context_mla_fp8_bf16_rope_dsa_fp8_index", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-awq-dsa-mla-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ/F16/F32 checkpoint bytes from safetensors headers. AWQ dequantization, sparse-attention kernel efficiency, speculative decode, and compute overhead are outside Bounds Engine v1.", "notes": "The config records AWQ 4-bit GEMM quantization with group size 128 and zero points. The bundled official inference config records dtype fp8 and the bundled code states deployment uses FP8 KV cache, so this profile reuses the audited V3.2 compressed-state cache coefficient." }, "evidence": [ { "label": "QuantTrio DeepSeek V3.2 AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/DeepSeek-V3.2-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 340023cb6036c97c5c664ac944300e9d2b1a3f2e, the API records a public MIT text-generation repo with transformers, safetensors, deepseek_v32, vLLM, AWQ, 4-bit, awq, endpoints_compatible, region:us, and base_model deepseek-ai/DeepSeek-V3.2. Current downloads were 431262 when audited. API safetensors metadata records F32 1075456, I32 681408069632, F16 3946184704, and logical total 685355329792. The model card points to the official vLLM DeepSeek V3.2 guide and gives a vLLM command with tensor parallelism and optional self-speculative decoding." }, { "label": "QuantTrio DeepSeek V3.2 AWQ config", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/raw/340023cb6036c97c5c664ac944300e9d2b1a3f2e/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_state", "serving" ], "notes": "The config records DeepseekV32ForCausalLM, bfloat16 model dtype, 61 hidden layers, 1 next-token-prediction layer, 3 initial dense layers, hidden size 7168, intermediate size 18432, MoE intermediate size 2048, 256 routed experts, 8 experts per token, 1 shared expert, 128 attention heads, 128 KV heads, q_lora_rank 1536, kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, index_head_dim 128, index_n_heads 64, index_topk 2048, 163840 max position embeddings, vocab size 129280, and AWQ 4-bit GEMM quantization with group size 128, zero points, and mlp.gate excluded from conversion." }, { "label": "DeepSeek V3.2 base config comparison", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/a7e62ac04ecb2c0a54d736dc46601c5606cf10a6/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences across 25 checked architecture fields between the AWQ config and the official DeepSeek V3.2 config. The checked fields covered architecture, dtype, layer counts, next-token-prediction layer count, dense/MoE transition, hidden dimensions, expert counts, attention geometry, MLA ranks, DSA index fields, context length, vocab size, and tied embeddings. The base config uses FP8 quantization metadata while the target uses AWQ metadata." }, { "label": "QuantTrio DeepSeek V3.2 AWQ safetensors headers", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/resolve/340023cb6036c97c5c664ac944300e9d2b1a3f2e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 121 indexed shards. Tensor payload spans sum exactly to the index total_size, 361.909457408 GB, across 138357 tensors: I32 343.365785088 GB, F16 18.539370496 GB, and F32 0.004301824 GB. Main tensors, defined as layers 0-60 plus embeddings, norm, and lm_head, sum to 352.008032512 GB. The auxiliary next-token-prediction layer 61 sums to 9.901424896 GB resident-only. Main routed expert tensors for layers 3-60 sum to 339.726041088 GB, exactly 1.327054848 GB per expert index. Main fixed ordinary-decode traffic sums to 12.281991424 GB." }, { "label": "QuantTrio DeepSeek V3.2 AWQ bundled official inference code", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/raw/340023cb6036c97c5c664ac944300e9d2b1a3f2e/inference/model.py", "source_type": "manual_review", "supports": [ "compressed_state", "kv_store_format", "kv_read_format", "dsa_index_cache" ], "notes": "The target repo's bundled inference/model.py is byte-identical to the official DeepSeek V3.2 file reviewed for the base profile. Manual review found MLA caches for kv_lora_rank 512 and qk_rope_head_dim 64, plus an Indexer cache with index_head_dim 128 stored as FP8 and F32 scale values. The code states actual deployment uses FP8 KV cache and uses index_topk to select sparse attention positions." }, { "label": "QuantTrio DeepSeek V3.2 AWQ bundled official inference config", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/raw/340023cb6036c97c5c664ac944300e9d2b1a3f2e/inference/config_671B_v3.2.json", "source_type": "config", "supports": [ "compressed_state", "serving" ], "notes": "The target repo's bundled inference/config_671B_v3.2.json is byte-identical to the official DeepSeek V3.2 file reviewed for the base profile. It records dtype fp8, scale_fmt ue8m0, 61 layers, kv_lora_rank 512, qk_rope_head_dim 64, index_head_dim 128, index_n_heads 64, and index_topk 2048." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card text, served AWQ config, base config comparison, bundled inference code/config comparison, direct safetensors header byte grouping, and the existing DeepSeek V3.2 official inference implementation review." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It is still audited so larger hardware can produce profile-backed bounds without falling back to the stale generated full-KV estimate." }, { "id": "quanttrio--gemma-4-31b-it-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/gemma-4-31B-it-AWQ", "title": "QuantTrio Gemma 4 31B IT AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ package of Gemma 4 31B IT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The repo records google/gemma-4-31B-it as its quantized base model. Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, dense setting, and vision geometry between this AWQ artifact and the Google BF16 base repo. The AWQ artifact adds a quantization_config with group_size 64 and leaves the vision tower plus language layer 0 unconverted." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 31.273088876, "swept_params_b": 30.69734534, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 20.459796184, "swept_weight_gb": 19.308309112, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "hf_api_logical_awq_parameters_with_exact_header_stored_bytes", "swept_parameter_scope": "model.language_model safetensors headers, including layer 0 unquantized BF16 tensors, AWQ qweight/qzeros tensors, F16 scales, norms, and tied embedding/output projection", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary generated text tokens", "notes": "HF API logical parameters are used for parameter counts, while exact range-read safetensors header bytes drive memory footprint and per-token traffic. The served config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6542301038801934, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored AWQ qweight/qzeros tensors, F16 scales, unquantized BF16 modules, and BF16 vision tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records AWQ GEMM 4-bit weights with group_size 64, zero_point true, and no KV-cache quantization scheme, so KV is charged at two bytes per scalar like the base BF16 Gemma profiles." }, "evidence": [ { "label": "QuantTrio Gemma 4 31B IT AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/gemma-4-31B-it-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 200d06ee83eb49a03c6e3120dbf7b09191eb1539, the API records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-31B-it, AWQ and vLLM tags, endpoints_compatible and region:us tags, 501734 downloads, and safetensors logical parameters BF16: 2465298284, I32: 28807790592, total: 31273088876." }, { "label": "QuantTrio Gemma 4 31B IT AWQ config", "url": "https://huggingface.co/QuantTrio/gemma-4-31B-it-AWQ/raw/200d06ee83eb49a03c6e3120dbf7b09191eb1539/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, AWQ 4-bit GEMM quantization, group_size 64, zero_point true, modules_to_not_convert vision_tower and model.layers.0, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 262144 max position embeddings, and a resident vision tower." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this QuantTrio AWQ artifact; the QuantTrio artifact adds quantization_config while preserving the base architecture." }, { "label": "QuantTrio Gemma 4 31B IT AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/gemma-4-31B-it-AWQ/raw/200d06ee83eb49a03c6e3120dbf7b09191eb1539/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records total_size 20459796184 bytes across seven shards. Range-read safetensors headers found 1994 tensors totaling 20.459796184 GB: 4.930596568 GB BF16 tensors, 0.900243456 GB F16 tensors, and 14.628956160 GB I32 tensors. Language tensors under model.language_model total 19.308309112 GB and are swept for ordinary text decode, including the unquantized layer 0 package and tied embedding/output projection. Resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.151487072 GB. The index has no separate lm_head.weight." }, { "label": "Google Gemma 4 31B IT BF16 profile", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_pattern", "base_model_proof" ], "notes": "This profile reuses the audited Gemma 4 31B text-decode KV layout from the Google BF16 profile because the QuantTrio config preserves the same 60-layer hybrid full/sliding attention architecture and does not enable KV-cache quantization." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current Hugging Face API metadata, model card, pinned served config, base config comparison, safetensors index, and direct shard-header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate. It is separate from the cyankiwi AWQ profile because the QuantTrio artifact uses a different AWQ format, group size, shard layout, and exact stored byte footprint." }, { "id": "quanttrio--qwen3-5-27b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3.5-27B-AWQ", "title": "QuantTrio Qwen3.5 27B AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3.5 27B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ artifact records Qwen/Qwen3.5-27B as its quantized base model. Manual comparison found the same audited top-level, text, layer-pattern, and vision geometry as the BF16 base: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The target changes served dtype fields to float16 and adds AWQ quantization metadata, so weight and KV precision are taken from the served target config." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 21.85285424, "swept_weight_gb": 17.539198464, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "base logical Qwen3.5 27B parameters with direct AWQ safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the package mixes AWQ-packed I32 qweight/qzeros tensors, F16 scale tensors, unquantized BF16 tensors, and tiny F32 tensors. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes. The AWQ module exclusions leave visual, MTP, layer 0, linear-attention in_proj_a/in_proj_b, and self-attention q/k/v tensors unquantized." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 F16 bytes))", "notes": "The AWQ artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The served text config dtype is float16, so the convolution state is charged as F16 while the recurrent matrix state is charged as F32. Read traffic charges one full fixed-state read per generated token; compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7865993885467936, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-awq-gemm-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ safetensors bytes: packed I32 qweight/qzeros tensors, F16 scale tensors, unquantized BF16 tensors, and F32 state parameters from shard headers. Dequantization, activation traffic, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group_size 128 and zero_point true, served text dtype float16, vision dtype float16, and no quantized KV cache scheme. This profile charges full-attention KV as FP16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "QuantTrio Qwen3.5 27B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-27B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 56f41874389615226dcd849ded92261a0286ff59, the API records a public Apache-2.0 image-text-to-text repo with transformers, safetensors, qwen3_5, vLLM, AWQ, 4-bit, awq, endpoints_compatible, region:us, and base_model Qwen/Qwen3.5-27B. Current downloads were 213771 when audited. The API safetensors block reports F32 8448, BF16 5011591664, I32 22769827840, and total 27781427952 storage-accounting tensor elements." }, { "label": "QuantTrio Qwen3.5 27B AWQ model card", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ/raw/56f41874389615226dcd849ded92261a0286ff59/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "architecture" ], "notes": "The card says this repo quantizes Qwen/Qwen3.5-27B using data-free quantization with no calibration dataset required. It gives a vLLM startup example using max_model_len 32768, tensor_parallel_size 2, and qwen3_next_mtp speculative decoding, while this profile models ordinary non-speculative text decode." }, { "label": "QuantTrio Qwen3.5 27B AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ/raw/56f41874389615226dcd849ded92261a0286ff59/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, AWQ 4-bit weights with group_size 128, zero_point true, GEMM version, served text dtype float16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, a resident float16 vision config, and one MTP layer. The modules_to_not_convert list covers visual, linear-attention in_proj_a/in_proj_b, self-attention q/k/v, layer 0, and MTP tensors." }, { "label": "Qwen3.5 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited top-level, text_config, layer_types, and vision_config geometry fields between the BF16 base config and this AWQ artifact after excluding quantization_config, name_or_path, dtype, and Transformers bookkeeping fields." }, { "label": "QuantTrio Qwen3.5 27B AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ/resolve/56f41874389615226dcd849ded92261a0286ff59/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split" ], "notes": "The index records total_size 21.852854240 GB across eight shards. Safetensors headers were range-read across all eight shards and found 1891 tensors totaling the same 21.852854240 GB: I32 11.473858560 GB, BF16 10.023183328 GB, F16 0.355778560 GB, and F32 0.000033792 GB. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 17.539198464 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 4.313655776 GB. Header buckets are language MLP 9.286471680 GB, language linear-attention tensors 3.098293760 GB, lm_head 2.542796800 GB, input embedding 2.542796800 GB, self-attention tensors 2.610315264 GB, visual 0.921460192 GB, MTP 0.849398784 GB, and language layer/norm tensors 0.001320960 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served AWQ config, pinned base config comparison, safetensors index, direct safetensors shard-header range reads, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and missed AWQ side tensors, unquantized modules, the resident-only multimodal/MTP/input-embedding split, and the fixed DeltaNet runtime state." }, { "id": "quanttrio--qwen3-5-27b-claude-4-6-opus-reasoning-distilled-v2-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ", "title": "QuantTrio Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled v2 AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of the Jackrong Qwen3.5 27B Claude 4.6 Opus reasoning-distilled v2 model.", "model_family": "qwen3.5-dense-multimodal-awq", "base_model_proof": { "base_model": "Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2", "relation": "quantized", "source": "Hugging Face model card/API metadata, served target config, public Jackrong base config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The AWQ artifact records Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 as its quantized base model. Manual comparison found matching checked top-level, text, layer-pattern, and vision geometry fields between the public Jackrong BF16 base config and this AWQ target config. The target adds AWQ quantization metadata and changes served top-level dtype fields to float16, so weight and KV precision are taken from the served target config and headers." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 21.85285424, "swept_weight_gb": 17.539198464, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "base logical Qwen3.5 27B parameters with direct AWQ safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from this repo's safetensors headers because the package mixes AWQ-packed I32 qweight/qzeros tensors, F16 scale tensors, unquantized BF16 tensors, and tiny F32 tensors. Logical parameter counts follow the audited Qwen3.5 27B architecture so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes. The AWQ module exclusions leave visual, MTP, layer 0, linear-attention in_proj_a/in_proj_b, and self-attention q/k/v tensors unquantized." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 F16 bytes))", "notes": "The AWQ artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The served target config dtype is float16, so the convolution state is charged as F16 while the recurrent matrix state is charged as F32. Read traffic charges one full fixed-state read per generated token; compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7865993885467936, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-awq-gemm-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ safetensors bytes: packed I32 qweight/qzeros tensors, F16 scale tensors, unquantized BF16 tensors, and F32 state parameters from shard headers. Dequantization, activation traffic, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group_size 128 and zero_point true, served text dtype float16, vision dtype float16, and no quantized KV cache scheme. This profile charges full-attention KV as FP16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "QuantTrio Qwen3.5 27B Claude v2 AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA c633586811ef0e1d5938809364d68f8a5e401fe0, the API records a public Apache-2.0 image-text-to-text repo with transformers, safetensors, qwen3_5, vLLM, AWQ, 4-bit, awq, endpoints_compatible, region:us, and base_model Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2. Current downloads were 216646 when audited. The API safetensors block reports F32 8448, BF16 5011591664, I32 22769827840, and total 27781427952 storage-accounting tensor elements." }, { "label": "QuantTrio Qwen3.5 27B Claude v2 AWQ model card", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ/raw/c633586811ef0e1d5938809364d68f8a5e401fe0/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "architecture" ], "notes": "The card says this repo quantizes the Jackrong v2 reasoning-distilled model using data-free quantization with no calibration dataset required. It gives a vLLM startup example using max_model_len 32768, tensor_parallel_size 2, and qwen3_next_mtp speculative decoding, while this profile models ordinary non-speculative text decode." }, { "label": "QuantTrio Qwen3.5 27B Claude v2 AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ/raw/c633586811ef0e1d5938809364d68f8a5e401fe0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, AWQ 4-bit weights with group_size 128, zero_point true, GEMM version, served top-level dtype float16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, a resident float16 vision config, and one MTP layer. The modules_to_not_convert list covers visual, linear-attention in_proj_a/in_proj_b, self-attention q/k/v, layer 0, and MTP tensors." }, { "label": "Jackrong Qwen3.5 27B Claude v2 base API and config comparison", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "The base repo API records an Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-27B with BF16/F32 safetensors total 27781427952. Manual comparison of checked architecture fields found no differences between the Jackrong BF16 base config and this AWQ target config for model class, model type, tied embeddings, text layer count, hidden size, intermediate size, attention/KV heads, head dimension, layer_types, linear-attention state geometry, MTP settings, vocab size, vision depth, vision hidden size, vision heads, and patch/merge settings." }, { "label": "QuantTrio Qwen3.5 27B Claude v2 AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ/resolve/c633586811ef0e1d5938809364d68f8a5e401fe0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split" ], "notes": "The index records total_size 21.852854240 GB across eight shards. Safetensors headers were range-read across all eight shards and found 1891 tensors totaling the same 21.852854240 GB: I32 11.473858560 GB, BF16 10.023183328 GB, F16 0.355778560 GB, and F32 0.000033792 GB. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 17.539198464 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 4.313655776 GB. Header buckets are language layers 14.996391424 GB, language final norm 0.000010240 GB, lm_head 2.542796800 GB, input embedding 2.542796800 GB, visual 0.921460192 GB, and MTP 0.849398784 GB. Linked-object HEAD checks resolved the eight shards to 21.853092832 GB total, leaving 238592 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served AWQ config, public Jackrong base config comparison, safetensors index, direct safetensors shard-header range reads, linked-object HEAD checks, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and missed AWQ side tensors, unquantized modules, the resident-only multimodal/MTP/input-embedding split, and the fixed DeltaNet runtime state. It is byte-identical to the audited QuantTrio Qwen3.5 27B AWQ package but has separate base-model evidence and current repo metadata." }, { "id": "quanttrio--qwen3-5-9b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3.5-9B-AWQ", "title": "QuantTrio Qwen3.5 9B AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served AWQ config, base config comparison, and direct safetensors shard header grouping", "config_compatible": true, "notes": "The artifact records Qwen/Qwen3.5-9B as its quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The AWQ artifact changes serving precision metadata while preserving the base hybrid attention geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 12.37637424, "swept_weight_gb": 8.943534592, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct AWQ safetensors stored-byte totals", "swept_parameter_scope": "model.language_model layers, final norm, and lm_head safetensors headers, excluding embed_tokens", "auxiliary_scope": "model.visual tensors, top-level MTP tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, BF16 unquantized embeddings, lm_head, vision, MTP, self-attention, linear-attention, first-layer MLP, and tiny F32 recurrent scalars. Logical parameter counts follow the audited Qwen3.5 BF16 architecture so model identity remains the 9.653104368B logical model while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The AWQ artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, dequantization, and activation traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The served config records no KV cache quantization, so KV cache is charged at two bytes per scalar." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.282113377021762, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ safetensors bytes: packed I32 qweights/qzeros, F16 scales, BF16 unquantized tensors, and tiny F32 recurrent scalar tensors from shard headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records text dtype float16 and AWQ GEMM 4-bit quantization with group_size 128, zero_point true, and modules_to_not_convert for visual, linear_attn, self_attn, model.layers.0, and MTP. Exact resident and swept byte fields drive production bounds." }, "evidence": [ { "label": "QuantTrio Qwen3.5 9B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-9B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 938f8e3ef86c9d1e9bec3705e149694c172592f1, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B, with vLLM, AWQ, 4-bit, awq, endpoints_compatible, and region:us tags. Current downloads are 453471. The API safetensors block reports F32: 3840, I32: 4680843264, BF16: 4972257264, and total: 9653104368 storage-accounting elements. The model card states this is a data-free AWQ quantization and gives a vLLM serving command with max model length 262144 and speculative MTP." }, { "label": "QuantTrio Qwen3.5 9B AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-9B-AWQ/raw/938f8e3ef86c9d1e9bec3705e149694c172592f1/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, AWQ GEMM 4-bit quantization with group_size 128 and zero_point true, modules_to_not_convert for visual, linear_attn, self_attn, model.layers.0, and MTP, text_config dtype float16, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching top-level identity fields and text geometry fields between the base BF16 repo and this AWQ artifact. The AWQ config uses float16 serving dtype labels and adds quantization metadata while preserving the hybrid full-attention/linear-attention layer pattern and one MTP layer." }, { "label": "QuantTrio Qwen3.5 9B AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3.5-9B-AWQ/raw/938f8e3ef86c9d1e9bec3705e149694c172592f1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 12376374240 bytes across 961 tensors and five shards. Direct range-read safetensors headers match that payload total: 12.376374240 GB, split into BF16 9.944514528 GB, I32 2.358706176 GB, F16 0.073138176 GB, and F32 0.000015360 GB. Linked-object HEAD checks for all five shards resolved to commit 938f8e3ef86c9d1e9bec3705e149694c172592f1, with combined linked file size 12.376491776 GB and 117536 bytes of safetensors header/container overhead outside tensor payloads. The swept ordinary text subset, defined as language layers plus final norm plus lm_head, excluding embed_tokens, totals 8.943534592 GB. The resident-only subset, defined as embed_tokens plus model.visual plus model.mtp, totals 3.432839648 GB. Header buckets are language linear-attention tensors 3.235398144 GB, language MLP tensors 2.733834240 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, self-attention tensors 0.939532288 GB, visual 0.912020960 GB, MTP 0.486581248 GB, language layer/norm tensors 0.000532480 GB, and language final norm 0.000008192 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served AWQ config, current base config comparison, safetensors index, linked-object HEAD checks, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and missed the AWQ zero/scale tensors plus unquantized embeddings, output head, self-attention, linear-attention, first language layer, MTP, and visual tensors." }, { "id": "quanttrio--qwen3-6-27b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3.6-27B-AWQ", "title": "QuantTrio Qwen3.6 27B AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ artifact records Qwen/Qwen3.6-27B as its quantized base model. Manual comparison found matching audited top-level, text, and vision geometry fields: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The AWQ artifact adds quantization_config and changes text/vision dtype labels to float16." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 21.852837344, "swept_weight_gb": 17.539181568, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "logical AWQ qweight plus BF16 model parameters represented by safetensors headers", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, and unquantized BF16 tensors. Logical parameter counts follow the Hugging Face API convention for this repo: I32 qweight tensors are counted as unpacked 4-bit logical parameters, BF16 tensors are counted by element, and qzeros/F16 scales are storage overhead rather than logical model parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The AWQ artifact preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled. The model card's vLLM example uses --max-model-len 32768 as a deployment cap, while the served config and base model card record 262144 native context." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, qzeros, F16 scales, and unquantized BF16 tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group_size 128, zero_point true, and modules_to_not_convert for visual, selected linear-attention inputs, self-attention Q/K/V projections, the first layer, and MTP. The text_config dtype is float16; KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "QuantTrio Qwen3.6 27B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.6-27B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 9b507bdc9afafb87b7898700cc2a591aa6639461, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-27B, with vLLM, AWQ, 4-bit, endpoints_compatible, and region:us tags. Current downloads are 1353250. The API safetensors block reports BF16: 5011600112, I32: 22769827840, total: 27781427952. The card says the repo was quantized using data-free quantization, identifies Qwen/Qwen3.6-27B as the base model, and gives a vLLM AWQ startup command." }, { "label": "QuantTrio Qwen3.6 27B AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ/raw/9b507bdc9afafb87b7898700cc2a591aa6639461/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, AWQ GEMM 4-bit quantization with group size 128 and zero points, text_config dtype float16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited top-level, text_config, layer_types, and vision_config geometry fields between the current base BF16 repo and this AWQ artifact. The AWQ artifact adds quantization_config and dtype labels but preserves the architecture fields used by this profile." }, { "label": "QuantTrio Qwen3.6 27B AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ/resolve/9b507bdc9afafb87b7898700cc2a591aa6639461/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 21852837344 bytes across eight shards and 1891 tensors. Range-read shard headers match the index exactly. Stored tensor bytes total 21.852837344 GB: BF16 10.023200224 GB, I32 11.473858560 GB, and F16 0.355778560 GB. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 17.539181568 GB and 25.624600064 logical parameters. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 4.313655776 GB and 2.156827888 logical parameters. Stored swept suffix bytes are qweight 11.384913920 GB, qzeros 0.088944640 GB, scales 0.355778560 GB, and unquantized BF16 swept tensors 5.709544448 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the model card, pinned AWQ config, current base config comparison, safetensors index, direct range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the AWQ package as ideal 4-bit dense weights and undercounted AWQ qzeros/scales plus unquantized attention, MTP, visual, embedding, and output-head tensors." }, { "id": "quanttrio--qwen3-6-35b-a3b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3.6-35B-A3B-AWQ", "title": "QuantTrio Qwen3.6 35B A3B AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ artifact records Qwen/Qwen3.6-35B-A3B as its quantized base model. Manual comparison found the same text and vision geometry as the audited base: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 experts per token, 1 shared expert, one MTP layer, and 262144 native context. The target changes serving dtype fields and adds AWQ quantization metadata, so weight and KV precision are taken from the served target config." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 25.44858288, "main_resident_weight_gb": 21.849040128, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.921536256, "routed_expert_weight_gb": 0.070029312, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_f16_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model layers, final norm, and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept as full matrices for each ordinary generated text token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are not routed expert tensors in the header grouping, so their traffic is included in fixed_weight_gb.", "notes": "Header-derived bytes are used because the package mixes AWQ I32 packed tensors, F16 side tensors, BF16 tensors, and F32 gates. Routed expert tensors are byte-uniform across 256 expert indexes after summing all 40 layers; layer 0 is stored larger than layers 1-39 but each expert index has the same total routed byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 F16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The served AWQ text config records dtype float16 and mamba_ssm_dtype float32. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, qzeros, scales, and unquantized tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group size 128 and zero points. Full-attention KV uses the served float16 dtype; DeltaNet recurrent state uses F32 for the recurrent matrix state and F16 for the convolution state." }, "evidence": [ { "label": "QuantTrio Qwen3.6 35B A3B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.6-35B-A3B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit 119886a1072372348f73ef0df2d801cdcc0f455b, the API records a public Apache-2.0 image-text-to-text repo with transformers, qwen3_5_moe, vLLM, AWQ, 4-bit, awq, endpoints_compatible, region:us, and base_model Qwen/Qwen3.6-35B-A3B. Current downloads were 695,633 when audited. The model card says this repo quantizes the model using a data-free quantization tool and no calibration dataset." }, { "label": "QuantTrio Qwen3.6 35B A3B AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.6-35B-A3B-AWQ/raw/119886a1072372348f73ef0df2d801cdcc0f455b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, AWQ quantization with group size 128, zero points, GEMM version, and a 40-layer text config with dtype float16, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, and one MTP layer." }, { "label": "Qwen3.6 35B A3B base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The base card states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the same audited text_config and vision_config geometry fields between the base BF16 repo and this AWQ artifact. The only text_config difference was dtype, and the only vision_config difference was torch_dtype. The AWQ artifact adds quantization_config while preserving the base architecture." }, { "label": "QuantTrio Qwen3.6 35B A3B AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3.6-35B-A3B-AWQ/resolve/119886a1072372348f73ef0df2d801cdcc0f455b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The index records total_size 25.448582880 GB across nine shards. Safetensors headers were range-read across all nine shards and found 92,355 tensors totaling the same 25.448582880 GB: 15.826157568 GB I32 tensors, 9.047805664 GB BF16 tensors, 0.490733568 GB F16 tensors, and 0.083886080 GB F32 tensors. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.849040128 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Routed expert tensors sum to 17.927503872 GB, or 0.070029312 GB per expert index. Fixed ordinary text traffic sums to 3.921536256 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, served AWQ config, base config comparison, safetensors index, direct shard-header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate for this AWQ repo, including the generated shared-expert count, resident weight, active weight, and missing fixed DeltaNet runtime state." }, { "id": "quanttrio--qwen3-coder-30b-a3b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ", "title": "QuantTrio Qwen3-Coder 30B A3B Instruct AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3-Coder 30B A3B Instruct.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served QuantTrio config, official BF16 base config comparison, and safetensors header review", "config_compatible": false, "notes": "The repo metadata records Qwen/Qwen3-Coder-30B-A3B-Instruct as its quantized base and preserves the core ordinary-decode topology: Qwen3MoeForCausalLM, 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 experts per token, no shared expert, 262144 max positions, and disabled sliding-window attention. Manual comparison against the audited official BF16 base config found real served-config differences in intermediate_size, max_window_layers, and explicit qk/qkv fields, so this profile uses the QuantTrio served config and tensor headers directly rather than claiming full base config compatibility." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b-awq-served", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.80267264, "main_resident_weight_gb": 16.180342784, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.11859712, "routed_expert_weight_gb": 0.117669888, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_bf16_f16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The config records shared_expert_intermediate_size 0. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived bytes are used because the AWQ package stores packed I32 qweight/qzeros tensors plus F16/BF16 side tensors. Routed expert tensors are stored as per-expert down_proj, gate_proj, and up_proj qweight/qzeros/scales tensors across all 48 layers; all 128 expert indexes are byte-uniform." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records sliding_window null and use_sliding_window false, so this profile charges full-context BF16 K and V streams for all 48 layers." }, "notes": "Qwen3MoeForCausalLM AWQ profile using the served QuantTrio config and direct safetensors header grouping. The profile models ordinary cached text decode; prefill, AWQ dequantization, activation traffic, and expert-parallel communication are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-qwen3-coder-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, qzeros, scales, and unquantized BF16/F16 tensors from safetensors headers. AWQ dequantization, activation traffic, kernel choice, and compute overhead are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group size 128, zero points, and modules_to_not_convert [\".mlp.gate\"]. The config does not define a KV-cache quantization scheme, so KV traffic is charged as BF16 from the served runtime dtype." }, "evidence": [ { "label": "QuantTrio Qwen3-Coder 30B A3B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit c58857a7f41c0920f73d1b56678640f9c02017d7, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, qwen3_moe, AWQ, 4-bit, awq, vLLM, endpoints_compatible, region:us, and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct tags. Current downloads are 267003. The API safetensors summary records I32: 29896998912, BF16: 635123712, total: 30532122624." }, { "label": "QuantTrio Qwen3-Coder 30B A3B Instruct AWQ model card", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ/raw/c58857a7f41c0920f73d1b56678640f9c02017d7/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The model card identifies the official Qwen3-Coder 30B A3B Instruct base, warns that this checkpoint has significant loss under 4-bit quantization, and gives a vLLM startup command requiring expert parallelism." }, { "label": "QuantTrio Qwen3-Coder 30B A3B Instruct AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ/raw/c58857a7f41c0920f73d1b56678640f9c02017d7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 5472, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, shared_expert_intermediate_size 0, decoder_sparse_step 1, max_position_embeddings 262144, max_window_layers 28, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 10000000, and AWQ quantization with 4 bits, group size 128, zero point true, GEMM version, and modules_to_not_convert [\".mlp.gate\"]." }, { "label": "Qwen3-Coder 30B A3B Instruct BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison checked 26 architecture fields. The QuantTrio config matches the official BF16 base on the MoE routing and KV-bearing decode fields used by this profile, including layer count, heads, KV heads, head_dim, expert count, experts per token, max positions, sliding_window null, and use_sliding_window false. It differs on intermediate_size 5472 vs 6144, max_window_layers 28 vs 48, and explicit qk/qkv fields, so config_compatible is false and fixed_weight_gb is derived from the QuantTrio tensor headers." }, { "label": "QuantTrio Qwen3-Coder 30B A3B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ/resolve/c58857a7f41c0920f73d1b56678640f9c02017d7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 16802672640 bytes across six shards. Direct range-read safetensors headers found 56115 tensors with payload bytes exactly matching the index total: 16.802672640 GB, split into I32 15.065284608 GB, BF16 1.270247424 GB, and F16 0.467140608 GB. Linked shard sizes total 16.809457184 GB, leaving 0.006784544 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately and remains in fixed decode traffic. Routed expert tensors sum to 15.061745664 GB and divide exactly into 128 uniform expert groups of 0.117669888 GB. Non-expert fixed decode tensors including lm_head.weight sum to 1.118597120 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, pinned served AWQ config, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads." }, "notes": "This self-contained profile supersedes the generated 0.5-byte metadata estimate for this AWQ repo. It uses exact stored AWQ tensor payload bytes and keeps KV as BF16 full-context because the served config does not request KV quantization." }, { "id": "quanttrio--qwen3-vl-30b-a3b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ", "title": "QuantTrio Qwen3 VL 30B A3B Instruct AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3-VL 30B A3B Instruct.", "model_family": "qwen3-vl-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ artifact records Qwen/Qwen3-VL-30B-A3B-Instruct as its base model and preserves the audited Qwen3-VL text and vision geometry: 48 text layers, 128 routed experts, 8 experts per token, no shared expert, 4 KV heads, 128 head dimension, and 262144 max position embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-vl-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.879935456, "main_resident_weight_gb": 16.180342784, "auxiliary_resident_weight_gb": 1.699592672, "fixed_weight_gb": 1.11859712, "routed_expert_weight_gb": 0.117669888, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_awq_i32_f16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding and visual tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records no shared expert. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived bytes are used because the package stores AWQ I32 packed tensors and F16 side tensors. Routed expert tensors are stored per expert index as qweight, qzeros, and scales tensors for down_proj, gate_proj, and up_proj across all 48 layers; all 128 expert indexes are byte-uniform." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 48 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-qwen3-vl-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored AWQ packed weights, qzeros, scales, and unquantized tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records AWQ GEMM 4-bit quantization with group size 128 and zero points. KV traffic is charged from the preserved base Qwen3-VL bfloat16 text geometry." }, "evidence": [ { "label": "QuantTrio Qwen3 VL 30B A3B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At commit a5ea10732e2c9330957864ece3ec66d806a4b00d, the API records an Apache-2.0 text-generation repo derived from Qwen/Qwen3-VL-30B-A3B-Instruct with AWQ/4-bit/vLLM tags, 1398983 downloads, used storage 17.887601176 GB, and safetensors parameters split across I32: 29896998912 and F16: 1173755120 tensors, total: 31070754032." }, { "label": "QuantTrio Qwen3 VL 30B A3B Instruct AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ/raw/a5ea10732e2c9330957864ece3ec66d806a4b00d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, AWQ quantization with group size 128, zero points, GEMM version, modules_to_not_convert visual and mlp.gate, a 48-layer text config, 32 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, 262144 max position embeddings, tie_word_embeddings false, and a resident 27-layer visual tower." }, { "label": "Qwen3 VL 30B A3B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo and this AWQ artifact; the AWQ artifact adds quantization_config while preserving the base architecture." }, { "label": "QuantTrio Qwen3 VL 30B A3B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ/resolve/a5ea10732e2c9330957864ece3ec66d806a4b00d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The index records total_size 17879935456 bytes across six shards. Safetensors headers were range-read across all six shards and found 56466 tensors totaling 17.879935456 GB: 15.065284608 GB I32 tensors and 2.814650848 GB F16 tensors. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 16.180342784 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1.699592672 GB. Main routed expert tensors sum to 15.061745664 GB, or 0.117669888 GB per expert index. Fixed ordinary text traffic sums to 1.11859712 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from model card/API metadata, served config, base config comparison, quantization config, and safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this AWQ repo, including exact stored AWQ side tensor overhead and resident-only multimodal package tensors." }, { "id": "quanttrio--qwen3-vl-32b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "QuantTrio/Qwen3-VL-32B-Instruct-AWQ", "title": "QuantTrio Qwen3 VL 32B Instruct AWQ", "summary": "Audited memory-side text-decode bounds profile for the QuantTrio AWQ 4-bit package of Qwen3-VL 32B Instruct.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, served config comparison, BF16 base profile, and safetensors header review", "config_compatible": true, "notes": "The model card and API metadata identify Qwen/Qwen3-VL-32B-Instruct as the quantized base. Manual comparison found no differences in checked top-level, text_config, or vision_config geometry fields between this AWQ repo, the audited BF16 base repo, and the official FP8 sibling; this repo only adds AWQ quantization metadata while preserving the served text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-32b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.357390064, "swept_params_b": 31.984210944, "auxiliary_resident_params_b": 1.37317912, "resident_weight_gb": 20.515832288, "swept_weight_gb": 17.769474048, "auxiliary_resident_weight_gb": 2.74635824, "resident_parameter_scope": "safetensors_header_awq_int4_i32_f16_tensor_payloads", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "Range-read safetensors headers record 1954 tensors totaling 20.515832288 GB, matching the index total_size. The swept subset includes AWQ qweight/qzeros/scales language layer tensors, F16 language norms, and the separate F16 lm_head.weight output projection. The root config records tie_word_embeddings false, so model.language_model.embed_tokens.weight is resident-only for ordinary decode while lm_head.weight remains swept." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 64 layers, 8 KV heads, and a 128 head dimension. The served config does not define sliding-window attention, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6150310995146205, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-f16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored AWQ safetensors bytes: I32 qweight/qzeros tensors plus F16 scales, visual tensors, embeddings, norms, and lm_head. AWQ dequantization, activation traffic, compute throughput, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records quant_method awq, bits 4, group_size 128, version gemm, zero_point true, modules_to_not_convert visual, and top-level torch_dtype float16. The model card recommends vLLM." }, "evidence": [ { "label": "QuantTrio Qwen3 VL 32B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-VL-32B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit c2d44b1376c73fa120fdb4e216c2ef61c1945498, the current API reports a public non-gated Apache-2.0 image-text-to-text Transformers repo with safetensors, qwen3_vl, AWQ, vLLM, endpoints_compatible, 4-bit, awq, region:us, base_model Qwen/Qwen3-VL-32B-Instruct, base_model_relation quantized, and 259994 downloads. The API safetensors block records I32: 31205621760, F16: 2151768304, total: 33357390064." }, { "label": "QuantTrio Qwen3 VL 32B AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-32B-Instruct-AWQ/raw/c2d44b1376c73fa120fdb4e216c2ef61c1945498/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLForConditionalGeneration, root tie_word_embeddings false, top-level torch_dtype float16, 64 text layers, hidden size 5120, intermediate size 25600, 64 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and AWQ quantization with 4 bits, group size 128, GEMM version, zero_point true, and visual modules excluded from conversion." }, { "label": "Qwen3 VL 32B Instruct BF16 base config and profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct/raw/0cfaf48183f594c314753d30a4c4974bc75f3ccb/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The already audited BF16 base profile records the same full-context text KV geometry used here: 64 text layers, hidden size 5120, 8 KV heads, 128 head dimension, 262144 context, untied embeddings, and a 27-layer visual tower. Manual config comparison found no checked geometry differences between the BF16 base, official FP8 sibling, and this AWQ artifact." }, { "label": "QuantTrio Qwen3 VL 32B AWQ safetensors index and shard headers", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-32B-Instruct-AWQ/raw/c2d44b1376c73fa120fdb4e216c2ef61c1945498/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "Range-read headers across all seven safetensors shards found 1954 tensors totaling 20.515832288 GB, matching index total_size: I32 payloads total 15.724707840 GB and F16 payloads total 4.791124448 GB. model.language_model.embed_tokens.weight has shape [151936, 5120] and contributes 0.777912320B parameters / 1.555824640 GB resident-only for ordinary decode. lm_head.weight is separate F16 with the same shape and remains in swept decode traffic. model.visual tensors total 0.595266800B parameters / 1.190533600 GB. Ordinary text swept traffic is 17.769474048 GB, and auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, total 1.373179120B parameters / 2.746358240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served AWQ config, audited BF16 base profile/config, official FP8 sibling config comparison, local scrape row, linked-object HEAD checks, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the generated ideal 0.5-byte estimate with exact AWQ safetensors payload bytes and preserves BF16 full-context KV because no quantized KV scheme is recorded." }, { "id": "qwen--qwen1-5-0-5b-chat", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen1.5-0.5B-Chat", "title": "Qwen1.5 0.5B Chat BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen1.5 0.5B Chat repo.", "model_family": "qwen1.5-dense", "architecture": { "canonical_architecture_id": "qwen1-5-0-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.619570176, "swept_params_b": 0.463987712, "auxiliary_resident_params_b": 0.155582464, "resident_weight_gb": 1.239140352, "swept_weight_gb": 0.927975424, "auxiliary_resident_weight_gb": 0.311164928, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers, model.norm, and lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary cached text decode token", "notes": "The config records tie_word_embeddings true, but the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic; the separate lm_head.weight remains swept for logits." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records num_key_value_heads 16, hidden_size 1024, 16 attention heads, sliding_window 32768, and use_sliding_window false, so Bounds Engine v1 charges BF16 full-context K/V cache traffic for all 24 layers." }, "notes": "Dense Qwen2ForCausalLM profile for the Qwen1.5 beta chat checkpoint, using the served repo config directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, attention kernels, cache writes, and tokenizer/chat-template overhead are outside this memory-side bound.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen1.5 0.5B Chat model card", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat", "source_type": "model_card", "supports": [ "repo", "license", "model_family", "max_context_tokens" ], "notes": "The card identifies this as Qwen1.5-0.5B-Chat under the Tongyi Qianwen research license. It describes Qwen1.5 as the beta version of Qwen2, says all sizes support 32K context, and notes that this beta release temporarily did not include GQA except for 32B or the mixture of sliding-window and full attention." }, { "label": "Qwen1.5 0.5B Chat config", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat/raw/4d14e384a4b037942bb3f3016665157c8bcb70ea/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, model_type qwen2, bfloat16, hidden_size 1024, intermediate_size 2816, 24 layers, 16 attention heads, 16 KV heads, max_position_embeddings 32768, sliding_window 32768, use_sliding_window false, max_window_layers 21, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen1.5 0.5B Chat Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-0.5B-Chat", "source_type": "derived_calculation", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 4d14e384a4b037942bb3f3016665157c8bcb70ea, the live API records a public non-gated text-generation repo with qwen2, chat, conversational, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 179717. The API safetensors block records BF16 619570176 parameters." }, { "label": "Qwen1.5 0.5B Chat safetensors header", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat/resolve/4d14e384a4b037942bb3f3016665157c8bcb70ea/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "dtype_split" ], "notes": "A range-read of model.safetensors found a 32992-byte header with 291 BF16 tensors totaling 619570176 parameters / 1.239140352 GB. model.embed_tokens.weight has shape [151936, 1024] and contributes 155582464 parameters / 0.311164928 GB. lm_head.weight is a separate BF16 tensor with the same shape and remains in swept decode traffic. Layer tensors plus model.norm and lm_head total 463987712 parameters / 0.927975424 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, model card, pinned config, tokenizer/chat config, and direct safetensors header grouping." }, "notes": "This profile replaces the rounded generated dense estimate with exact BF16 stored bytes and separates the resident-only input embedding from ordinary text-decode traffic." }, { "id": "qwen--qwen1-5-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen1.5-7B", "title": "Qwen1.5 7B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen1.5 7B base repo.", "model_family": "qwen1.5-dense", "architecture": { "canonical_architecture_id": "qwen1-5-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.721324544, "swept_params_b": 7.098994688, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 15.442649088, "swept_weight_gb": 14.197989376, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers, model.norm, and lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary cached text decode token", "notes": "The config records tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic; the separate lm_head.weight remains swept for logits." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records num_key_value_heads 32, hidden_size 4096, 32 attention heads, sliding_window 32768, and use_sliding_window false, so Bounds Engine v1 charges BF16 full-context K/V cache traffic for all 32 layers." }, "notes": "Dense Qwen2ForCausalLM profile for the Qwen1.5 beta base checkpoint, using the served repo config directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, attention kernels, cache writes, and tokenizer overhead are outside this memory-side bound.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen1.5 7B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-7B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 831096e3a59a0789a541415da25ef195ceb802fe, the live API records a public non-gated text-generation repo with transformers, safetensors, qwen2, pretrained, conversational, text-generation-inference, endpoints_compatible, deploy:azure, region:us, and license other / tongyi-qianwen tags. Current downloads are 111122. The API safetensors block records BF16 7721324544 parameters. The model card describes Qwen1.5 as the beta version of Qwen2, says the 7B size is one of the dense releases, and says all sizes support 32K context." }, { "label": "Qwen1.5 7B config", "url": "https://huggingface.co/Qwen/Qwen1.5-7B/raw/831096e3a59a0789a541415da25ef195ceb802fe/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, model_type qwen2, bfloat16, hidden_size 4096, intermediate_size 11008, 32 layers, 32 attention heads, 32 KV heads, max_position_embeddings 32768, sliding_window 32768, use_sliding_window false, max_window_layers 28, vocab_size 151936, tie_word_embeddings false, and rope_theta 1000000." }, { "label": "Qwen1.5 7B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen1.5-7B/raw/831096e3a59a0789a541415da25ef195ceb802fe/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "dtype_split" ], "notes": "The index lists four safetensors shards and records total_size 15442649088 bytes. Range-read shard headers found 387 BF16 tensors totaling 7721324544 parameters / 15.442649088 GB, matching index total_size. Linked-object range checks resolved combined linked file size 15.442693552 GB with 44464 bytes of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 4096] and contributes 622329856 parameters / 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7098994688 parameters / 14.197989376 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, model card, pinned config, and direct safetensors shard-header grouping." }, "notes": "This profile replaces the rounded generated dense estimate with exact BF16 stored bytes and separates the resident-only input embedding from ordinary text-decode traffic." }, { "id": "qwen--qwen1-5-moe-a2-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen1.5-MoE-A2.7B", "title": "Qwen1.5 MoE A2.7B BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 Qwen1.5 MoE A2.7B base repo.", "model_family": "qwen2-moe", "architecture": { "canonical_architecture_id": "qwen1-5-moe-a2-7b", "max_context_tokens": 8192, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 28.631568384, "main_resident_weight_gb": 28.009238528, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 3.095072768, "routed_expert_weight_gb": 0.415236096, "routed_experts": 60, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers.0-23, model.norm, and lm_head, with per-expert tensor groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "Qwen2MoeSparseMoeBlock always computes shared_expert and shared_expert_gate before adding them to routed expert output. Shared expert tensors are therefore included in fixed_weight_gb.", "notes": "All 24 layers are sparse MoE layers because decoder_sparse_step is 1 and mlp_only_layers is empty. Routed experts are stored as per-expert gate/up/down matrices under model.layers.*.mlp.experts.*, so routed_expert_weight_gb is the grouped expert tensor byte count divided by 60 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The pinned Transformers Qwen2-MoE implementation updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache for all 24 layers." }, "notes": "This profile targets the base pretraining repo directly. It uses exact stored BF16 tensor bytes and treats the input embedding as resident-only for ordinary decode, while lm_head remains swept fixed traffic because embeddings are untied." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-qwen2-moe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Bounds Engine v1 charges stored BF16 safetensors bytes, BF16 K/V cache traffic, and exact expert groups; activation traffic, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The config records torch_dtype bfloat16, and the API safetensors block plus direct shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen1.5 MoE A2.7B API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-MoE-A2.7B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 1a758c50ecb6350748b9ce0a99d2352fd9fc11c9, the live API records a public non-gated text-generation repo with transformers, safetensors, qwen2_moe, pretrained, moe, endpoints_compatible, license:other, and region:us tags. Current downloads are 216320. The API safetensors block records BF16 14315784192 and total 14315784192." }, { "label": "Qwen1.5 MoE A2.7B model card", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/raw/1a758c50ecb6350748b9ce0a99d2352fd9fc11c9/README.md", "source_type": "model_card", "supports": [ "repo", "license", "architecture", "total_params_b", "active_params_b" ], "notes": "The model card describes Qwen1.5-MoE-A2.7B as a transformer decoder-only MoE base model upcycled from Qwen-1.8B, with 14.3B total parameters and about 2.7B activated parameters during runtime." }, { "label": "Qwen1.5 MoE A2.7B config", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/raw/1a758c50ecb6350748b9ce0a99d2352fd9fc11c9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen2MoeForCausalLM, qwen2_moe, bfloat16, 24 layers, hidden size 2048, 16 attention heads, 16 KV heads, derived head_dim 128, 60 experts, 4 experts per token, decoder_sparse_step 1, shared_expert_intermediate_size 5632, use_sliding_window false, tie_word_embeddings false, vocab size 151936, and max_position_embeddings 8192." }, { "label": "Transformers Qwen2-MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/0b05eeced8e8e4a5b2682000a8c1b151e76a9181/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py", "source_type": "manual_review", "supports": [ "routed_experts", "shared_experts_per_token", "kv_adapter" ], "notes": "Manual review found Qwen2MoeSparseMoeBlock computes a top-k routed expert path plus shared_expert and shared_expert_gate, then adds shared output to the expert output. Qwen2MoeAttention updates standard key_states and value_states in past_key_values, matching the full-context expanded K/V cache adapter when use_sliding_window is false." }, { "label": "Qwen1.5 MoE A2.7B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/resolve/1a758c50ecb6350748b9ce0a99d2352fd9fc11c9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 28631568384 bytes across 8 referenced files. Direct range-read safetensors headers found 4659 BF16 tensors with payload bytes 28.631568384 GB and header/container overhead 0.000576560 GB. model.embed_tokens.weight contributes 0.622329856 GB resident-only. Ordinary text resident bytes therefore sum to 28.009238528 GB. Routed expert tensors sum to 24.914165760 GB and divide into 60 uniform expert indexes of 0.415236096 GB. Fixed ordinary text traffic, including self-attention, routers, shared experts, shared expert gates, norms, and lm_head, sums to 3.095072768 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, served config, pinned Transformers implementation, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated estimate with exact BF16 stored bytes and explicitly models the always-on shared expert path." }, { "id": "qwen--qwen2-0-5b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-0.5B-Instruct", "title": "Qwen2 0.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2 0.5B Instruct repo.", "model_family": "qwen2-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-0.5B", "relation": "finetune", "source": "Hugging Face model metadata, served Instruct config, base config comparison, and direct safetensors header grouping", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings while the base repo records 131072, and the EOS token differs. This profile therefore uses the served Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-0-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 0.494032768, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.136134656, "non_embedding_params_b": 0.357898112, "notes": "A range-read of model.safetensors records 494032768 BF16 stored parameters. model.embed_tokens.weight contributes 136134656 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 2, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 64 head dimension is derived from hidden_size 896 divided by 14 attention heads." }, "notes": "Dense Qwen2ForCausalLM instruction-tuned profile using the served repo config rather than copying the base model's larger context field." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2 0.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "training_stage" ], "notes": "The card identifies this as the instruction-tuned 0.5B Qwen2 model, Apache-2.0 licensed, with base_model Qwen/Qwen2-0.5B. It describes the Qwen2 decoder architecture with SwiGLU, attention QKV bias, and grouped-query attention." }, { "label": "Qwen2 0.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct/raw/c540970f9e29518b1d8f06ab8b24cba66ad77b6d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 896, intermediate_size 4864, 24 layers, 14 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, use_sliding_window false, max_window_layers 24, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen2 0.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-0.5B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "repo", "pipeline", "downloads", "base_model_proof" ], "notes": "At commit c540970f9e29518b1d8f06ab8b24cba66ad77b6d, the API records downloads 555044, public text-generation metadata, Apache-2.0 licensing, region:us, base_model Qwen/Qwen2-0.5B, and safetensors parameters BF16: 494032768, total: 494032768." }, { "label": "Qwen2 0.5B base config", "url": "https://huggingface.co/Qwen/Qwen2-0.5B/raw/91d2aff3f957f99e4c74c962f2f408dcc88a18d8/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different max_position_embeddings and EOS fields, so this profile does not copy the base config wholesale." }, { "label": "Qwen2 0.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct/resolve/c540970f9e29518b1d8f06ab8b24cba66ad77b6d/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 32280-byte header with 290 BF16 tensors totaling 494032768 parameters / 0.988065536 GB. model.embed_tokens.weight has shape [151936, 896] and contributes 136134656 parameters / 0.272269312 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, model card, base config comparison, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for the instruction-tuned repo, separate from the Qwen2 0.5B base profile because the served context field differs." }, { "id": "qwen--qwen2-0-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-0.5B", "title": "Qwen2 0.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2 0.5B base repo.", "model_family": "qwen2-dense", "architecture": { "canonical_architecture_id": "qwen2-0-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense", "total_params_b": 0.494032768, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.136134656, "non_embedding_params_b": 0.357898112, "notes": "A range-read of model.safetensors records 494032768 BF16 stored parameters. model.embed_tokens.weight contributes 136134656 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 2, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 64 head dimension is derived from hidden_size 896 divided by 14 attention heads." }, "notes": "Dense Qwen2ForCausalLM base-pretraining profile using the served repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2 0.5B model card", "url": "https://huggingface.co/Qwen/Qwen2-0.5B", "source_type": "model_card", "supports": [ "repo", "license", "training_stage", "newer_version" ], "notes": "The card identifies this as the base 0.5B Qwen2 language model, Apache-2.0 licensed, and points to Qwen/Qwen2.5-0.5B as the newer version. It describes the Qwen2 decoder architecture with SwiGLU, attention QKV bias, and grouped-query attention." }, { "label": "Qwen2 0.5B config", "url": "https://huggingface.co/Qwen/Qwen2-0.5B/raw/91d2aff3f957f99e4c74c962f2f408dcc88a18d8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 896, intermediate_size 4864, 24 layers, 14 attention heads, 2 KV heads, 131072 max position embeddings, sliding_window 131072, use_sliding_window false, max_window_layers 24, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen2 0.5B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-0.5B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "repo", "pipeline", "downloads" ], "notes": "At commit 91d2aff3f957f99e4c74c962f2f408dcc88a18d8, the API records downloads 1017191, public text-generation metadata, Apache-2.0 licensing, region:us, and safetensors parameters BF16: 494032768, total: 494032768." }, { "label": "Qwen2 0.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2-0.5B/resolve/91d2aff3f957f99e4c74c962f2f408dcc88a18d8/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 32280-byte header with 290 BF16 tensors totaling 494032768 parameters / 0.988065536 GB. model.embed_tokens.weight has shape [151936, 896] and contributes 136134656 parameters / 0.272269312 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from the served config, model card/API metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for the base pretrained repo, separate from the newer Qwen2.5 0.5B profile." }, { "id": "qwen--qwen2-1-5b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-1.5B-Instruct", "title": "Qwen2 1.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2 1.5B Instruct repo.", "model_family": "qwen2-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-1.5B", "relation": "finetune", "source": "Model card description and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records eos_token_id 151645, 32768 max position embeddings, and sliding_window 32768 while the base config records eos_token_id 151643, 131072 max position embeddings, and sliding_window 131072. This profile uses the served Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters. tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2 1.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The card identifies this as the instruction-tuned 1.5B Qwen2 model, describes the Qwen2 release as base plus aligned chat models, and records Apache-2.0 licensing." }, { "label": "Qwen2 1.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct/raw/ba1cf1846d7df0a0591d6c00649f57e798519da8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 1536, 12 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, use_sliding_window false, and tie_word_embeddings true." }, { "label": "Qwen2 1.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-1.5B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit ba1cf1846d7df0a0591d6c00649f57e798519da8, the API safetensors block records BF16: 1543714304 and total: 1543714304, which this profile stores as 1.543714304B parameters." }, { "label": "Qwen2 1.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct/resolve/ba1cf1846d7df0a0591d6c00649f57e798519da8/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2-1.5B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different context and eos fields, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-1-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-1.5B", "title": "Qwen2 1.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2 1.5B base repo.", "model_family": "qwen2-dense", "architecture": { "canonical_architecture_id": "qwen2-1-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters. tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM base-pretraining profile using the served base repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2 1.5B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-1.5B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 8a16abf2848eda07cc5253dec660bf1ce007ad7a, the API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen2, pretrained, conversational, en, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags; 197198 downloads; and safetensors parameters BF16: 1543714304, total: 1543714304. The model card identifies this as the 1.5B Qwen2 base language model." }, { "label": "Qwen2 1.5B config", "url": "https://huggingface.co/Qwen/Qwen2-1.5B/raw/8a16abf2848eda07cc5253dec660bf1ce007ad7a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 1536, intermediate_size 8960, 28 layers, 12 attention heads, 2 KV heads, 131072 max position embeddings, sliding_window 131072, use_sliding_window false, max_window_layers 28, vocab_size 151936, eos_token_id 151643, and rope_theta 1000000." }, { "label": "Qwen2 1.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2-1.5B/resolve/8a16abf2848eda07cc5253dec660bf1ce007ad7a/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2 1.5B Instruct config comparison", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct/raw/ba1cf1846d7df0a0591d6c00649f57e798519da8/config.json", "source_type": "config", "supports": [ "architecture" ], "notes": "Manual comparison against the audited Qwen2 1.5B Instruct profile found matching core tensor geometry, dtype, tokenizer size, tied-embedding setting, and safetensors storage. The base config records eos_token_id 151643, max_position_embeddings 131072, and sliding_window 131072, while the Instruct config records eos_token_id 151645, max_position_embeddings 32768, and sliding_window 32768." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served config, direct safetensors header grouping, and comparison with the audited Qwen2 1.5B Instruct sibling." }, "notes": "This is a self-contained dense BF16 profile for the base pretrained repo, separate from the existing instruction-tuned profile." }, { "id": "qwen--qwen2-5-0-5b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-0.5B-Instruct-GGUF", "title": "Qwen2.5 0.5B Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Qwen2.5 0.5B Instruct.", "model_family": "qwen2.5-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-0.5B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen2.5 0.5B Instruct base profile", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2.5-0.5B-Instruct. The selected FP16 GGUF header records the same dense Qwen2 tensor geometry as the audited BF16 base profile: 24 layers, hidden size 896, intermediate size 4864, 14 attention heads, 2 KV heads, 64 key/value head dimension, and tied base-model embeddings represented as separate token_embd.weight and output.weight tensors in GGUF. The selected GGUF header records qwen2.context_length 8192, while the model card and BF16 base config record 32768 full context with 8192 generation, so this profile uses the selected artifact header as serving truth and records the mismatch explicitly." }, "architecture": { "canonical_architecture_id": "qwen2-5-0-5b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.630167424, "swept_params_b": 0.494032768, "auxiliary_resident_params_b": 0.136134656, "resident_weight_gb": 1.266425696, "swept_weight_gb": 0.98820864, "auxiliary_resident_weight_gb": 0.278217056, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for qwen2.5-0.5b-instruct-fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 1.266425696 GB. GGUF tensor spans total 1.260477952 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005947744 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 2, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 24 Qwen2 decoder layers, 2 KV heads, 64-dimensional key/value heads, and qwen2.context_length 8192. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The card lists multiple lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF. It stores most tensors as F16 and small norm/bias tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Qwen2.5 0.5B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-0.5B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 9217f5db79a29953eb74d5343926648285ec7e67, the API records a public non-gated Apache-2.0 GGUF text-generation repo with base_model Qwen/Qwen2.5-0.5B-Instruct, base_model:quantized metadata, endpoints_compatible, region:us, 148822 downloads, GGUF architecture qwen2, context_length 8192, gguf.total 630167424, and gguf.totalFileSize 1266425696." }, { "label": "Qwen2.5 0.5B Instruct GGUF model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "context_mismatch" ], "notes": "The card records Apache-2.0 licensing, original model Qwen/Qwen2.5-0.5B-Instruct, llama.cpp usage, quantized sibling files q2_K through q8_0, 0.49B parameters, 0.36B non-embedding parameters, 24 layers, 14 Q heads, 2 KV heads, 32768 full context, and 8192 generation. The quickstart example downloads the Q5_K_M file, but API gguf.totalFileSize selects the FP16 file for this repo-level profile." }, { "label": "Qwen2.5 0.5B Instruct audited BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records Qwen2ForCausalLM, bfloat16, 24 layers, hidden size 896, intermediate size 4864, 14 attention heads, 2 KV heads, 64 head dimension derived from hidden size over attention heads, 32768 max positions, tied embeddings, and use_sliding_window false. This GGUF profile reuses only the matching tensor geometry and keeps the selected GGUF header's 8192 context limit." }, { "label": "Qwen2.5 0.5B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/tree/9217f5db79a29953eb74d5343926648285ec7e67", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 1.266425696 GB, Q8_0 0.675710816 GB, Q6_K 0.650379104 GB, Q5_K_M 0.522186592 GB, Q5_0 0.490475360 GB, Q4_K_M 0.491400032 GB, Q4_0 0.428730208 GB, Q3_K_M 0.432041824 GB, and Q2_K 0.415182688 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Qwen2.5 0.5B Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/9217f5db79a29953eb74d5343926648285ec7e67/qwen2.5-0.5b-instruct-fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "An 8MB range-read of the GGUF v3 header found 26 metadata entries and 291 tensors. The linked file is 1.266425696 GB. Tensor spans sum to 1.260477952 GB: output.weight 0.272269312 GB, token_embd.weight 0.272269312 GB, blk.* tensors 0.715935744 GB, and output_norm.weight 0.000003584 GB. Metadata/tokenizer/header/file overhead accounts for 0.005947744 GB. Stored tensor bytes split into F16 1.260191744 GB and F32 0.000286208 GB. The header records general.architecture qwen2, qwen2.block_count 24, context_length 8192, embedding_length 896, feed_forward_length 4864, attention.head_count 14, attention.head_count_kv 2, and rope.freq_base 1000000." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked-object HEAD checks for all GGUF siblings, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Qwen2.5 0.5B Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or a 32768-token serving context from the card without selecting and auditing a specific sibling artifact." }, { "id": "qwen--qwen2-5-0-5b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-0.5B-Instruct", "title": "Qwen2.5 0.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 0.5B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-0.5B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry and max context match the base config, but the Instruct repo records different eos_token_id and max_window_layers fields. This profile uses the Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-0-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 0.494032768, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.136134656, "non_embedding_params_b": 0.357898112, "notes": "A range-read of model.safetensors records 494032768 BF16 stored parameters. model.embed_tokens.weight contributes 136134656 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 2, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 64 head dimension is derived from hidden_size 896 divided by 14 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 0.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "max_context_tokens" ], "notes": "The model metadata identifies Qwen/Qwen2.5-0.5B as the base model and Apache-2.0 as the license. The card describes full 32768-token context for this instruction-tuned 0.5B model." }, { "label": "Qwen2.5 0.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/raw/7ae557604adf67be50417f59c2c2f167def9a775/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, tie_word_embeddings true, bfloat16, 24 layers, 2 KV heads, hidden_size 896, 14 attention heads, 32768 max position embeddings, and use_sliding_window false." }, { "label": "Qwen2.5 0.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-0.5B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "base_model_proof" ], "notes": "At commit 7ae557604adf67be50417f59c2c2f167def9a775, the API safetensors block records BF16: 494032768 and total: 494032768, which this profile stores as 0.494032768B parameters." }, { "label": "Qwen2.5 0.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/7ae557604adf67be50417f59c2c2f167def9a775/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 32280-byte header with 290 BF16 tensors totaling 494032768 parameters / 0.988065536 GB. model.embed_tokens.weight has shape [151936, 896] and contributes 136134656 parameters / 0.272269312 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 0.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different eos_token_id and max_window_layers fields, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-0-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-0.5B", "title": "Qwen2.5 0.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 0.5B base repo.", "model_family": "qwen2.5-dense", "architecture": { "canonical_architecture_id": "qwen2-5-0-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 0.494032768, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.136134656, "non_embedding_params_b": 0.357898112, "notes": "A range-read of model.safetensors records 494032768 BF16 stored parameters. model.embed_tokens.weight contributes 136134656 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 2, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 64 head dimension is derived from hidden_size 896 divided by 14 attention heads." }, "notes": "Dense Qwen2ForCausalLM base-pretraining profile using the served repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 0.5B model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B", "source_type": "model_card", "supports": [ "repo", "license", "training_stage", "max_context_tokens" ], "notes": "The card identifies this as the base 0.5B Qwen2.5 model, Apache-2.0 licensed, with pretraining stage, RoPE, SwiGLU, RMSNorm, attention QKV bias, tied word embeddings, 0.49B total parameters, 0.36B non-embedding parameters, 24 layers, 14 query heads, 2 KV heads, and full 32768-token context." }, { "label": "Qwen2.5 0.5B config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B/raw/060db6499f32faf8b98477b0a26969ef7d8b9987/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 896, intermediate_size 4864, 24 layers, 14 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, use_sliding_window false, max_window_layers 24, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen2.5 0.5B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-0.5B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "repo", "pipeline" ], "notes": "At commit 060db6499f32faf8b98477b0a26969ef7d8b9987, the API records text-generation, transformers, safetensors, qwen2, Apache-2.0 licensing, and safetensors parameters BF16: 494032768, total: 494032768." }, { "label": "Qwen2.5 0.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B/resolve/060db6499f32faf8b98477b0a26969ef7d8b9987/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 32280-byte header with 290 BF16 tensors totaling 494032768 parameters / 0.988065536 GB. model.embed_tokens.weight has shape [151936, 896] and contributes 136134656 parameters / 0.272269312 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for the base pretrained repo, separate from the existing instruction-tuned profile." }, { "id": "qwen--qwen2-5-1-5b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-1.5B-Instruct-AWQ", "title": "Qwen2.5 1.5B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2.5 1.5B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, and served config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-1.5B-Instruct as its quantized base model. Manual comparison found no differences in audited architecture, context, tokenizer, and attention geometry fields between the AWQ config and the base Instruct config after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version." }, "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.777088, "swept_params_b": 1.543714304, "auxiliary_resident_params_b": 0.233373696, "resident_weight_gb": 1.614472192, "swept_weight_gb": 1.1477248, "auxiliary_resident_weight_gb": 0.466747392, "resident_parameter_scope": "logical Qwen2.5 1.5B Instruct parameters represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and the separately stored lm_head.weight tensor", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings; the AWQ file separately stores lm_head.weight even though the config records tie_word_embeddings true", "notes": "Bounds use exact stored bytes from the safetensors header because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases/norms, and unquantized F16 embedding/head tensors. Exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.9084931033240897, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases/norms, and unquantized F16 embedding/head tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar. weight_bytes_per_param records resident stored bytes divided by logical safetensors parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 1.5B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "At commit 3ecffa0ceb27851800f45519bab9c457a04405e1, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-1.5B-Instruct, text-generation-inference, endpoints_compatible, 4-bit, AWQ, and region:us tags. Current downloads are 863664. The API safetensors block reports logical parameters split across I32: 1310195712 and F16: 466892288, total 1777088000." }, { "label": "Qwen2.5 1.5B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-AWQ/raw/3ecffa0ceb27851800f45519bab9c457a04405e1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen2ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 28 layers, 2 KV heads, hidden_size 1536, 12 attention heads, intermediate_size 8960, 32768 max position embeddings, tie_word_embeddings true, and use_sliding_window false." }, { "label": "Qwen2.5 1.5B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/989aa7980e4cf806f80c7fef2b1adb7bc71aa306/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the base Instruct config after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version." }, { "label": "Qwen2.5 1.5B Instruct AWQ safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-AWQ/resolve/3ecffa0ceb27851800f45519bab9c457a04405e1/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The linked file size is 1.614553840 GB. A direct range-read found an 81640-byte safetensors header plus 731 tensors totaling 1.614472192 GB: 0.660215808 GB I32 tensors and 0.954256384 GB F16 tensors. model.embed_tokens.weight is F16 with shape [151936, 1536] and contributes 233373696 logical parameters / 0.466747392 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 1.147724800 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, served AWQ config, base Instruct config comparison, and a direct safetensors header range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, duplicate head storage, and unquantized embedding/head tensors." }, { "id": "qwen--qwen2-5-1-5b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-1.5B-Instruct-GGUF", "title": "Qwen2.5 1.5B Instruct GGUF FP16", "summary": "Audited memory-side text-decode bounds profile for the API-selected FP16 GGUF artifact of Qwen2.5 1.5B Instruct.", "model_family": "qwen2.5-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected GGUF header metadata, and audited BF16 Qwen2.5 1.5B Instruct base profile", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2.5-1.5B-Instruct. The selected FP16 GGUF header records the same dense Qwen2 tensor geometry as the audited BF16 base profile: 28 layers, hidden size 1536, intermediate size 8960, 12 attention heads, 2 KV heads, 128 key/value head dimension, and tied base-model embeddings represented as separate token_embd.weight and output.weight tensors in GGUF. The selected GGUF header records qwen2.context_length 8192, while the model card and BF16 base config record 32768 full context with 8192 generation, so this profile uses the selected artifact header as serving truth and records the mismatch explicitly." }, "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.777088, "swept_params_b": 1.543714304, "auxiliary_resident_params_b": 0.233373696, "resident_weight_gb": 3.560416288, "swept_weight_gb": 3.0877184, "auxiliary_resident_weight_gb": 0.472697888, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for qwen2.5-1.5b-instruct-fp16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected FP16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected FP16 linked file is 3.560416288 GB. GGUF tensor spans total 3.554465792 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005950496 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen2 decoder layers, 2 KV heads, 128-dimensional key/value heads, and qwen2.context_length 8192. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected FP16 GGUF artifact. The card lists multiple lower-bit quantized GGUF siblings; those should get separate profiles if selected." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected FP16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the FP16 GGUF. It stores most tensors as F16 and small norm/bias tensors as F32. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Qwen2.5 1.5B Instruct GGUF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 91cad51170dc346986eccefdc2dd33a9da36ead9, the API records a public non-gated Apache-2.0 GGUF text-generation repo with base_model Qwen/Qwen2.5-1.5B-Instruct, base_model:quantized metadata, endpoints_compatible, region:us, 234624 downloads, GGUF architecture qwen2, context_length 8192, gguf.total 1777088000, and gguf.totalFileSize 3560416288." }, { "label": "Qwen2.5 1.5B Instruct GGUF model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "context_mismatch" ], "notes": "The card records Apache-2.0 licensing, original model Qwen/Qwen2.5-1.5B-Instruct, llama.cpp usage, quantized sibling files q2_K through q8_0, 1.54B parameters, 1.31B non-embedding parameters, 28 layers, 12 Q heads, 2 KV heads, 32768 full context, and 8192 generation. The quickstart example downloads the Q5_K_M file, but API gguf.totalFileSize selects the FP16 file for this repo-level profile." }, { "label": "Qwen2.5 1.5B Instruct audited BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 1536, intermediate size 8960, 12 attention heads, 2 KV heads, 128 head dimension, 32768 max positions, tied embeddings, and use_sliding_window false. This GGUF profile reuses only the matching tensor geometry and keeps the selected GGUF header's 8192 context limit." }, { "label": "Qwen2.5 1.5B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/tree/91cad51170dc346986eccefdc2dd33a9da36ead9", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found FP16 3.560416288 GB, Q8_0 1.894532128 GB, Q6_K 1.464178720 GB, Q5_K_M 1.285494304 GB, Q5_0 1.259173408 GB, Q4_K_M 1.117320736 GB, Q4_0 1.066227232 GB, Q3_K_M 0.924455968 GB, and Q2_K 0.752880160 GB. The selected FP16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Qwen2.5 1.5B Instruct FP16 GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/resolve/91cad51170dc346986eccefdc2dd33a9da36ead9/qwen2.5-1.5b-instruct-fp16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "An 8MB range-read of the GGUF v3 header found 26 metadata entries and 339 tensors. The linked file is 3.560416288 GB. Tensor spans sum to 3.554465792 GB: output.weight 0.466747392 GB, token_embd.weight 0.466747392 GB, blk.* tensors 2.620964864 GB, and output_norm.weight 0.000006144 GB. Metadata/tokenizer/header/file overhead accounts for 0.005950496 GB. Stored tensor bytes split into F16 3.553886208 GB and F32 0.000579584 GB. The header records general.architecture qwen2, qwen2.block_count 28, context_length 8192, embedding_length 1536, feed_forward_length 8960, attention.head_count 12, attention.head_count_kv 2, and rope.freq_base 1000000." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, audited BF16 base profile, linked-object HEAD checks for all GGUF siblings, and a direct GGUF header/tensor-index range read of the selected FP16 artifact." }, "notes": "Use this profile for the API-selected Qwen2.5 1.5B Instruct FP16 GGUF artifact. Do not infer the lower-bit sibling footprints or a 32768-token serving context from the card without selecting and auditing a specific sibling artifact." }, { "id": "qwen--qwen2-5-1-5b-instruct", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-1.5B-Instruct", "title": "Qwen2.5 1.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 1.5B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings and sliding_window while the base config records 131072. This profile uses the Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 1.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen2.5-1.5B as the base model and Apache-2.0 as the license." }, { "label": "Qwen2.5 1.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen2ForCausalLM, bfloat16, 28 layers, 2 KV heads, hidden_size 1536, 12 attention heads, 32768 max position embeddings, and use_sliding_window false." }, { "label": "Qwen2.5 1.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format" ], "notes": "At commit 989aa7980e4cf806f80c7fef2b1adb7bc71aa306, the API safetensors block records BF16: 1543714304 and total: 1543714304, which this profile stores as 1.543714304B parameters." }, { "label": "Qwen2.5 1.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/resolve/989aa7980e4cf806f80c7fef2b1adb7bc71aa306/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different context-related fields, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-1-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-1.5B", "title": "Qwen2.5 1.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 1.5B base repo.", "model_family": "qwen2.5-dense", "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM base-pretraining profile using the served base repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 1.5B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit 8faed761d45a263340a0528343f099c05c9a4323, the API records an Apache-2.0 text-generation repo with transformers, safetensors, qwen2, eval-results, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags; 1209897 downloads; and safetensors parameters BF16: 1543714304, total: 1543714304." }, { "label": "Qwen2.5 1.5B config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 1536, intermediate_size 8960, 28 layers, 12 attention heads, 2 KV heads, 131072 max position embeddings, sliding_window 131072, use_sliding_window false, max_window_layers 28, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen2.5 1.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/resolve/8faed761d45a263340a0528343f099c05c9a4323/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for the base pretrained repo, separate from the existing instruction-tuned profile." }, { "id": "qwen--qwen2-5-14b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-14B-Instruct-AWQ", "title": "Qwen2.5 14B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2.5 14B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-14B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-14B-Instruct as its quantized base model. Manual comparison found matching Qwen2ForCausalLM tensor geometry and context settings; the AWQ artifact adds quantization_config and changes torch_dtype from bfloat16 to float16." }, "architecture": { "canonical_architecture_id": "qwen2-5-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 9.980028928, "swept_weight_gb": 8.422893568, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical Qwen2.5 14B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, F16 scales, F16 biases, and unquantized F16 embeddings/head tensors. Logical parameter counts match the BF16 base model; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 14B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-14B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 539535859b135b0244c91f3e59816150c8056698, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-14B-Instruct, 4-bit and AWQ tags, 1446368 downloads, and logical safetensors parameters split across I32: 13212057600 and F16: 1557976064, total 14770033664." }, { "label": "Qwen2.5 14B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ/raw/539535859b135b0244c91f3e59816150c8056698/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen2ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 48 layers, 8 KV heads, hidden_size 5120, 40 attention heads, 32768 max position embeddings, tie_word_embeddings false, and use_sliding_window false." }, { "label": "Qwen2.5 14B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the base Instruct config after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version." }, { "label": "Qwen2.5 14B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ/raw/539535859b135b0244c91f3e59816150c8056698/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The index records total_size 9980028928 bytes across three shards. Range-read safetensors headers found 1251 tensors totaling 9.980028928 GB: 6.6576384 GB I32 tensors and 3.322390528 GB F16 tensors. model.embed_tokens.weight is F16 with shape [152064, 5120] and contributes 778567680 logical parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 8.422893568 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served AWQ config, base Instruct config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, and unquantized embedding/head tensors." }, { "id": "qwen--qwen2-5-14b-instruct-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4", "title": "Qwen2.5 14B Instruct GPTQ Int4", "summary": "Audited memory-side bounds profile for the official Qwen2.5 14B Instruct GPTQ 4-bit checkpoint.", "model_family": "qwen2.5-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-14B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served GPTQ config, current base-model config metadata, and safetensors shard headers", "config_compatible": true, "notes": "The GPTQ repo card and API metadata record Qwen/Qwen2.5-14B-Instruct as the quantized base model. Manual comparison against the audited BF16 base config found matching checked architecture and context fields; the GPTQ repo changes torch_dtype to float16 and adds GPTQ 4-bit quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2-5-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 9.988581376, "swept_weight_gb": 8.431446016, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical GPTQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 qweight/qzeros/g_idx tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts match the BF16 base model and Hugging Face API total; exact resident and swept GB fields drive production bounds." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 48 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes full 131072-token context via YaRN configuration changes, which are outside this default artifact profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.676273433306103, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-gptq-exllama-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ packed weights, qzeros, g_idx tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. GPTQ dequantization, ExLlama kernels, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 14B Instruct GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 08fdbde5eea15cd2aae884125fbfd604dec19818, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen2.5-14B-Instruct, Apache-2.0 license, tags 4-bit and gptq, region:us, safetensors logical parameters I32: 13212057600 and F16: 1557976064, total: 14770033664, and current downloads 111341." }, { "label": "Qwen2.5 14B Instruct GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4/blob/08fdbde5eea15cd2aae884125fbfd604dec19818/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 4-bit instruction-tuned 14B Qwen2.5 model, records 48 layers, GQA with 40 Q heads and 8 KV heads, GPTQ 4-bit quantization, and notes that inputs beyond 32768 tokens require adding YaRN rope scaling." }, { "label": "Qwen2.5 14B Instruct GPTQ Int4 served config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4/raw/08fdbde5eea15cd2aae884125fbfd604dec19818/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 13824, 48 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true." }, { "label": "Qwen2.5 14B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture and context fields between the BF16 base config and the GPTQ artifact config after excluding torch_dtype, quantization_config, and transformers_version: Qwen2ForCausalLM, hidden_size 5120, intermediate_size 13824, 48 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, vocab_size 152064, rope_theta 1000000, and rms_norm_eps 1e-6." }, { "label": "Qwen2.5 14B Instruct GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4/resolve/08fdbde5eea15cd2aae884125fbfd604dec19818/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 9988581376 bytes across three shards, matching direct range-read shard header spans. Headers contain 1587 tensors totaling 9.988581376 GB: 6.666190848 GB I32 tensors and 3.322390528 GB F16 tensors. Stored suffix totals are qweight 6.606028800 GB, qzeros 0.051609600 GB, g_idx 0.008552448 GB, scales 0.206438400 GB, F16 bias tensors 0.000688128 GB, and F16 weight tensors 3.115264000 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB. Ordinary text swept traffic is total payload minus input embedding, or 8.431446016 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served GPTQ config, audited base config comparison, safetensors index metadata, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen2.5 14B Instruct GPTQ Int4 artifact. Do not substitute the AWQ profile; the geometry is the same but the GPTQ package carries additional g_idx storage and a distinct commit pin." }, { "id": "qwen--qwen2-5-14b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-14B-Instruct", "title": "Qwen2.5 14B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 14B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-14B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Manual comparison found matching tensor geometry and dtype between the Instruct and base configs, but max_position_embeddings, max_window_layers, rms_norm_eps, and eos_token_id differ. This profile therefore uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 29.540067328, "swept_weight_gb": 27.982931968, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 579 BF16 tensors totaling 14770033664 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes 128K long-context support via optional YaRN, but the current config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 14B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-14B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "architecture", "total_params_b", "weight_format" ], "notes": "At commit cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-14B and safetensors parameters BF16: 14770033664. The model card identifies this as the instruction-tuned 14B Qwen2.5 model, with 14.7B parameters, 13.1B non-embedding parameters, 48 layers, 40 Q heads, 8 KV heads, and optional YaRN long-context handling." }, { "label": "Qwen2.5 14B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 48 layers, hidden_size 5120, 40 attention heads, 8 KV heads, intermediate_size 13824, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, rms_norm_eps 1e-6, vocab_size 152064, and 32768 max position embeddings." }, { "label": "Qwen2.5 14B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 29540067328 bytes across eight shards. Range-read shard headers found 579 BF16 tensors totaling 14770033664 parameters / 29.540067328 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 5120] and contributes 778567680 parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 13991465984 parameters / 27.982931968 GB." }, { "label": "Qwen2.5 14B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B/raw/97e1e76335b7017d8f67c08a19d103c0504298c9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, hidden size, layer count, attention head count, KV head count, intermediate size, dtype, vocab size, sliding_window, use_sliding_window, and tied-embedding setting. The base config differs on eos_token_id, max_position_embeddings, max_window_layers, and rms_norm_eps, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, base config comparison, safetensors index, direct range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-32b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-32B-Instruct-AWQ", "title": "Qwen2.5 32B Instruct AWQ", "summary": "Audited memory-side bounds profile for the official Qwen2.5 32B Instruct AWQ 4-bit checkpoint.", "model_family": "qwen2.5-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served AWQ config, current base-model API/config metadata, and safetensors shard headers", "config_compatible": true, "notes": "The AWQ repo card and API metadata record Qwen/Qwen2.5-32B-Instruct as the quantized base model. The served AWQ config preserves the current base config's Qwen2ForCausalLM geometry while adding AWQ GEMM 4-bit quantization." }, "architecture": { "canonical_architecture_id": "qwen2.5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 19.328804864, "swept_weight_gb": 17.771669504, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros, scales, and biases are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes full 131072-token context via YaRN configuration changes, which are outside this default artifact profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5900474986395108, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 32B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen2.5-32B-Instruct, Apache-2.0 license, tags 4-bit and awq, region:us, safetensors logical parameters I32: 31205621760 and F16: 1558254592, total: 32763876352, and current downloads 787926." }, { "label": "Qwen2.5 32B Instruct AWQ served config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/raw/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Qwen2.5 32B Instruct AWQ model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/blob/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the AWQ-quantized 4-bit instruction-tuned 32B Qwen2.5 model, records 32.5B parameters, 64 layers, GQA with 40 Q heads and 8 KV heads, AWQ 4-bit quantization, and notes that the current config is set for 32768 tokens while long context requires adding YaRN rope scaling." }, { "label": "Qwen2.5 32B Instruct base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "The live base-model API response at commit 5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd records a public Transformers Qwen2 text-generation repo, Apache-2.0 license, BF16 safetensors total 32763876352 parameters, and base_model Qwen/Qwen2.5-32B." }, { "label": "Qwen2.5 32B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture fields between the current base config and the AWQ artifact config: Qwen2ForCausalLM, hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, vocab_size 152064, and rope_theta 1000000." }, { "label": "Qwen2.5 32B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/raw/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 19328804864 bytes across five shards, matching direct range-read shard header spans. Headers contain 1667 tensors totaling 19.328804864 GB: 15.724707840 GB I32 tensors and 3.604097024 GB F16 tensors. Stored suffix totals are qweight 15.602810880 GB, qzeros 0.121896960 GB, scales 0.487587840 GB, F16 bias tensors 0.000917504 GB, and F16 weight tensors 3.115591680 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB." }, { "label": "Qwen2.5 32B Instruct AWQ linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/tree/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for all five safetensors shards resolved to x-repo-commit 5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c with linked sizes 3943575336, 3980134208, 3947411392, 3980134240, and 3477738728 bytes. The linked file sizes include safetensors JSON header overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served AWQ config, model card, current base-model API/config comparison, safetensors index metadata, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen2.5 32B Instruct AWQ artifact. Do not substitute the Qwen2.5 Coder 32B AWQ profile even though the geometry and byte split match; each repo keeps its own evidence and commit pins." }, { "id": "qwen--qwen2-5-32b-instruct-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4", "title": "Qwen2.5 32B Instruct GPTQ Int4", "summary": "Audited memory-side bounds profile for the official Qwen2.5 32B Instruct GPTQ 4-bit checkpoint.", "model_family": "qwen2.5-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served GPTQ config, current base-model API/config metadata, and safetensors shard headers", "config_compatible": true, "notes": "The GPTQ repo card and API metadata record Qwen/Qwen2.5-32B-Instruct as the quantized base model. Manual comparison against the current base config found matching checked architecture fields; the GPTQ repo changes torch_dtype to float16 and adds GPTQ 4-bit quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2.5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 19.343747072, "swept_weight_gb": 17.786611712, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical GPTQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 qweight/qzeros/g_idx tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the current BF16 base profile and Hugging Face API total; exact resident/swept byte fields drive production bounds." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes full 131072-token context via YaRN configuration changes, which are outside this default artifact profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.590398610475137, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-gptq-exllama-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ packed weights, qzeros, g_idx tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. GPTQ dequantization, ExLlama kernels, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 32B Instruct GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit c83e67dfb2664f5039fd4cd99e206799e27dd800, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen2.5-32B-Instruct, Apache-2.0 license, tags 4-bit and gptq, region:us, safetensors logical parameters I32: 31205621760 and F16: 1558254592, total: 32763876352, and current downloads 463895." }, { "label": "Qwen2.5 32B Instruct GPTQ Int4 served config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4/raw/c83e67dfb2664f5039fd4cd99e206799e27dd800/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true." }, { "label": "Qwen2.5 32B Instruct GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4/blob/c83e67dfb2664f5039fd4cd99e206799e27dd800/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 4-bit instruction-tuned 32B Qwen2.5 model, records 64 layers, GQA with 40 Q heads and 8 KV heads, GPTQ 4-bit quantization, and notes that inputs beyond 32768 tokens require adding YaRN rope scaling." }, { "label": "Qwen2.5 32B Instruct base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "The live base-model API response at commit 5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd records a public Transformers Qwen2 text-generation repo, Apache-2.0 license, BF16 safetensors total 32763876352 parameters, and base_model Qwen/Qwen2.5-32B." }, { "label": "Qwen2.5 32B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture fields between the current base config and the GPTQ artifact config: Qwen2ForCausalLM, hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, vocab_size 152064, rope_theta 1000000, and rms_norm_eps 1e-6." }, { "label": "Qwen2.5 32B Instruct GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4/resolve/c83e67dfb2664f5039fd4cd99e206799e27dd800/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 19343747072 bytes across five shards, matching direct range-read shard header spans. Headers contain 2115 tensors totaling 19.343747072 GB: 15.739650048 GB I32 tensors and 3.604097024 GB F16 tensors. Stored suffix totals are qweight 15.602810880 GB, qzeros 0.121896960 GB, g_idx 0.014942208 GB, scales 0.487587840 GB, F16 bias tensors 0.000917504 GB, and F16 weight tensors 3.115591680 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served GPTQ config, model card, current base-model API/config comparison, safetensors index metadata, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen2.5 32B Instruct GPTQ Int4 artifact. Do not substitute the AWQ profile; the geometry is the same but the GPTQ package carries additional g_idx storage and a distinct commit pin." }, { "id": "qwen--qwen2-5-32b-instruct-gptq-int8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8", "title": "Qwen2.5 32B Instruct GPTQ Int8", "summary": "Audited memory-side bounds profile for the official Qwen2.5 32B Instruct GPTQ 8-bit checkpoint.", "model_family": "qwen2.5-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served GPTQ config, current base-model API/config metadata, and safetensors shard headers", "config_compatible": true, "notes": "The GPTQ repo card and API metadata record Qwen/Qwen2.5-32B-Instruct as the quantized base model. Manual comparison against the current base config found matching checked architecture fields; the GPTQ repo changes torch_dtype to float16 and adds GPTQ 8-bit quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2.5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 35.068454912, "swept_weight_gb": 33.511319552, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical GPTQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 qweight/qzeros/g_idx tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the current BF16 base profile and Hugging Face API total; exact resident/swept byte fields drive production bounds." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes full 131072-token context via YaRN configuration changes, which are outside this default artifact profile." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1.070339007974535, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-gptq-exllama-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ packed weights, qzeros, g_idx tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. GPTQ dequantization, ExLlama kernels, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 8-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 32B Instruct GPTQ Int8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen2.5-32B-Instruct, Apache-2.0 license, tags 8-bit and gptq, region:us, safetensors logical parameters I32: 31205621760 and F16: 1558254592, total: 32763876352, and current downloads 118191." }, { "label": "Qwen2.5 32B Instruct GPTQ Int8 served config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8/raw/eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and GPTQ 8-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true." }, { "label": "Qwen2.5 32B Instruct GPTQ Int8 model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8/blob/eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 8-bit instruction-tuned 32B Qwen2.5 model, records 64 layers, GQA with 40 Q heads and 8 KV heads, GPTQ 8-bit quantization, and notes that inputs beyond 32768 tokens require adding YaRN rope scaling." }, { "label": "Qwen2.5 32B Instruct base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "The live base-model API response at commit 5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd records a public Transformers Qwen2 text-generation repo, Apache-2.0 license, BF16 safetensors total 32763876352 parameters, and base_model Qwen/Qwen2.5-32B." }, { "label": "Qwen2.5 32B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture fields between the current base config and the GPTQ artifact config: Qwen2ForCausalLM, hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, vocab_size 152064, rope_theta 1000000, rms_norm_eps 1e-6, and eos_token_id 151645." }, { "label": "Qwen2.5 32B Instruct GPTQ Int8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8/resolve/eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 35068454912 bytes across nine shards, matching direct range-read shard header spans. Headers contain 2115 tensors totaling 35.068454912 GB: 31.464357888 GB I32 tensors and 3.604097024 GB F16 tensors. Stored suffix totals are qweight 31.205621760 GB, qzeros 0.243793920 GB, g_idx 0.014942208 GB, scales 0.487587840 GB, F16 bias tensors 0.000917504 GB, and F16 weight tensors 3.115591680 GB. Linked shard sizes total 35.068693560 GB, leaving 0.000238648 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, served GPTQ config, model card, current base-model API/config comparison, safetensors index metadata, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen2.5 32B Instruct GPTQ Int8 artifact. Do not substitute the Int4 or AWQ profiles; the geometry is the same but the GPTQ storage bytes and commit pin are distinct." }, { "id": "qwen--qwen2-5-32b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-32B-Instruct", "title": "Qwen2.5 32B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 32B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings while the base config records 131072. The served Instruct config also differs in rms_norm_eps and eos_token_id, so this profile uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 65.527752704, "swept_weight_gb": 63.970617344, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 32763876352 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes long-context support, but the current served config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 32B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit 5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-32B and safetensors parameters BF16: 32763876352." }, { "label": "Qwen2.5 32B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 64 layers, hidden_size 5120, 40 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, rms_norm_eps 1e-6, and 32768 max position embeddings." }, { "label": "Qwen2.5 32B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 65527752704 bytes across 17 shards. Range-read shard headers found 771 BF16 tensors totaling 32763876352 parameters / 65.527752704 GB, matching the index total. model.embed_tokens.weight has shape [152064, 5120] and contributes 778567680 parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 31.985308672 parameters / 63.970617344 GB." }, { "label": "Qwen2.5 32B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B/raw/1818d35814b8319459f4bd55ed1ac8709630f003/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different max_position_embeddings, rms_norm_eps, and eos_token_id values, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served config, base config comparison, safetensors index, range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-3b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-3B-Instruct-AWQ", "title": "Qwen2.5 3B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2.5 3B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served AWQ config, and base config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-3B-Instruct as its quantized base model. Manual comparison found matching Qwen2ForCausalLM geometry, context fields, sliding-window settings, vocabulary size, and tied-embedding setting between the AWQ and BF16 base configs after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version." }, "architecture": { "canonical_architecture_id": "qwen2-5-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.397103616, "swept_params_b": 3.085938688, "auxiliary_resident_params_b": 0.311164928, "resident_weight_gb": 2.686599168, "swept_weight_gb": 2.064269312, "auxiliary_resident_weight_gb": 0.622329856, "resident_parameter_scope": "logical Qwen2.5 3B Instruct parameters represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and the separately stored lm_head.weight tensor", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings; the AWQ file separately stores lm_head.weight even though the config records tie_word_embeddings true", "notes": "Bounds use exact stored bytes from the safetensors header because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases/norms, and unquantized F16 embedding/head tensors. Exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7908499332626774, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases/norms, and unquantized F16 embedding/head tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar. weight_bytes_per_param records resident stored bytes divided by logical safetensors parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 3B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "At commit 3559b226e8ce77211e2c1bd7ddfb7686fec4d6dd, the API records a public non-gated Qwen Research License text-generation repo with base_model Qwen/Qwen2.5-3B-Instruct, text-generation-inference, endpoints_compatible, 4-bit, AWQ, and region:us tags. Current downloads are 127257. The API safetensors block reports logical parameters split across I32: 2774532096 and F16: 622571520, total 3397103616." }, { "label": "Qwen2.5 3B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-AWQ/raw/3559b226e8ce77211e2c1bd7ddfb7686fec4d6dd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen2ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 36 layers, 2 KV heads, hidden_size 2048, 16 attention heads, intermediate_size 11008, 32768 max position embeddings, tie_word_embeddings true, and use_sliding_window false. The stale sliding_window and max_window_layers fields do not apply while use_sliding_window is false." }, { "label": "Qwen2.5 3B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/aa8e72537993ba99e69dfaafa59ed015b17504d1/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The base repo records Qwen2ForCausalLM, bfloat16, 36 layers, 2 KV heads, hidden_size 2048, 16 attention heads, intermediate_size 11008, 32768 max position embeddings, tie_word_embeddings true, use_sliding_window false, and BF16 safetensors total 3085938688 logical parameters. Manual comparison found matching audited geometry and context fields." }, { "label": "Qwen2.5 3B Instruct AWQ safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-AWQ/resolve/3559b226e8ce77211e2c1bd7ddfb7686fec4d6dd/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The linked file size is 2.686704624 GB. A direct range-read found a 105448-byte safetensors header plus 939 tensors totaling 2.686599168 GB: 1.398104064 GB I32 tensors and 1.288495104 GB F16 tensors. Payload groups are qweight 1.387266048 GB, qzeros 0.010838016 GB, scales 0.043352064 GB, bias 0.000184320 GB, model.layers.* 1.441935360 GB, model.norm.weight 0.000004096 GB, model.embed_tokens.weight 0.622329856 GB, and lm_head.weight 0.622329856 GB. model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 logical parameters / 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 2.064269312 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, served AWQ config, base Instruct config comparison, linked-object HEAD metadata, and a direct safetensors header range read." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, duplicate head storage, and unquantized embedding/head tensors." }, { "id": "qwen--qwen2-5-3b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-3B-Instruct-GGUF", "title": "Qwen2.5 3B Instruct GGUF Q5_K_M", "summary": "Audited memory-side text-decode bounds profile for the Qwen2.5 3B Instruct Q5_K_M GGUF artifact.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, audited BF16 base profile, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The official GGUF repo records Qwen/Qwen2.5-3B-Instruct as its quantized base. The selected GGUF header records the same Qwen2ForCausalLM text geometry as the audited BF16 Instruct profile: 36 layers, 16 attention heads, 2 KV heads, 128 head dimension, 32768-token context, and full-context attention." }, "architecture": { "canonical_architecture_id": "qwen2-5-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.397103616, "swept_params_b": 3.085938688, "auxiliary_resident_params_b": 0.311164928, "resident_weight_gb": 2.438740384, "swept_weight_gb": 2.218858496, "auxiliary_resident_weight_gb": 0.219881888, "resident_parameter_scope": "selected qwen2.5-3b-instruct-q5_k_m.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.35 tensors from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "The BF16 base profile is tied and has no separate lm_head tensor, but the selected Q5_K_M GGUF stores both token_embd.weight and output.weight. This profile charges output.weight as swept decode traffic and treats token_embd.weight as resident-only input embedding storage. The selected linked file is 2.438740384 GB. Header tensor spans total 2.432784384 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005956000 GB." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The audited BF16 base config and selected GGUF header record ordinary full-context Qwen2 attention: 36 layers, 2 KV heads, and 128 key/value head dimension." }, "notes": "This profile models ordinary text decode for the selected main Q5_K_M GGUF artifact." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7178881363858877, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-qwen2.5-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The model card's download example selects qwen2.5-3b-instruct-q5_k_m.gguf, while the live HF API gguf.totalFileSize matches qwen2.5-3b-instruct-q2_k.gguf. This profile intentionally targets the card-example Q5_K_M artifact." }, "evidence": [ { "label": "Qwen2.5 3B Instruct GGUF HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "The live HF API response at commit 7dabda4d13d513e3e842b20f0d435c732f172cbe records an official Qwen GGUF repo with base_model Qwen/Qwen2.5-3B-Instruct, 215418 downloads, region:us, GGUF architecture qwen2, 32768 context length, gguf.total 3397103616, and gguf.totalFileSize 1376856480. The API totalFileSize matches qwen2.5-3b-instruct-q2_k.gguf, while this profile targets Q5_K_M because the card's CLI download example selects that file." }, { "label": "Qwen2.5 3B Instruct GGUF model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/raw/7dabda4d13d513e3e842b20f0d435c732f172cbe/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "serving" ], "notes": "The pinned card records the Qwen research license, text-generation pipeline, base_model Qwen/Qwen2.5-3B-Instruct, Qwen2ForCausalLM-style architecture, tied word embeddings, 3.09B parameters, 36 layers, 16 query heads, 2 KV heads, 32768 context, and available GGUF quantizations. Its Hugging Face CLI example downloads qwen2.5-3b-instruct-q5_k_m.gguf." }, { "label": "Qwen2.5 3B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct", "source_type": "derived_calculation", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "The existing audited BF16 profile records Qwen2ForCausalLM, 36 layers, 2 KV heads, 128 head dimension, full-context attention, 32768 max positions, BF16 safetensors, and tied embeddings without a separate lm_head tensor." }, { "label": "Qwen2.5 3B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/tree/7dabda4d13d513e3e842b20f0d435c732f172cbe", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found q2_k 1.376856480 GB, q3_k_m 1.724178848 GB, q4_0 1.997879712 GB, q4_k_m 2.104932768 GB, q5_0 2.383591840 GB, q5_k_m 2.438740384 GB, q6_k 2.793410976 GB, q8_0 3.616088480 GB, and two FP16 shards totaling 6.800646784 GB." }, { "label": "Qwen2.5 3B Instruct Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/resolve/7dabda4d13d513e3e842b20f0d435c732f172cbe/qwen2.5-3b-instruct-q5_k_m.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 26 metadata entries and 435 tensors. The linked file is 2.438740384 GB. Tensor spans sum to 2.432784384 GB across 3.397103616B logical elements: output.weight 0.255252480 GB, output_norm.weight 0.000008192 GB, token_embd.weight 0.213925888 GB, and blk.0-35 tensors 1.963597824 GB. Metadata/tokenizer/header/file overhead accounts for 0.005956000 GB. Tensor spans split into Q5_K 1.835941888 GB, Q6_K 0.595875840 GB, and F32 0.000966656 GB. The header records qwen2.block_count 36, context_length 32768, embedding_length 2048, feed_forward_length 11008, attention.head_count 16, attention.head_count_kv 2, and rope.freq_base 1000000." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, existing BF16 base profile, linked GGUF file sizes, and a direct selected-GGUF header/tensor-index range read." }, "notes": "Use this profile for the official Qwen Q5_K_M GGUF text artifact in ordinary text-decode bounds. Other GGUF quantizations in this repo have different resident and traffic bytes and require separate workload selection." }, { "id": "qwen--qwen2-5-3b-instruct", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-3B-Instruct", "title": "Qwen2.5 3B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 3B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-3B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching tensor geometry, context fields, sliding-window settings, and dtype between the Instruct and base configs." }, "architecture": { "canonical_architecture_id": "qwen2-5-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 3.085938688, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.311164928, "non_embedding_params_b": 2.77477376, "notes": "Range-read safetensors headers record 3085938688 BF16 stored parameters. model.embed_tokens.weight contributes 311164928 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 3B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen2.5-3B as the base model. The scraped catalog records license:other for this repo." }, { "label": "Qwen2.5 3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen2ForCausalLM, bfloat16, 36 layers, 2 KV heads, hidden_size 2048, 16 attention heads, 32768 max position embeddings, and use_sliding_window false." }, { "label": "Qwen2.5 3B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format" ], "notes": "At commit aa8e72537993ba99e69dfaafa59ed015b17504d1, the API safetensors block records BF16: 3085938688 and total: 3085938688, which this profile stores as 3.085938688B parameters." }, { "label": "Qwen2.5 3B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/aa8e72537993ba99e69dfaafa59ed015b17504d1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "The index lists two safetensors shards. Range-read shard headers record 434 BF16 tensors totaling 3085938688 parameters and 6.171877376 GB, matching index total_size. model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 parameters / 0.622329856 GB. The index has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 3B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant architecture and context fields match the Instruct repo config." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-3B", "title": "Qwen2.5 3B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 3B base repo.", "model_family": "qwen2.5-dense", "architecture": { "canonical_architecture_id": "qwen2-5-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 3.085938688, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.311164928, "non_embedding_params_b": 2.77477376, "notes": "Range-read safetensors headers record 3085938688 BF16 stored parameters. model.embed_tokens.weight contributes 311164928 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Dense Qwen2ForCausalLM base-pretraining profile using the served base repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 3B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "architecture", "total_params_b", "weight_format" ], "notes": "At repo SHA 3aab1f1954e9cc14eb9509a215f9e5ca08227a9b, the API records a public non-gated qwen-research text-generation repo with safetensors, qwen2, eval-results, and region:us tags. Current downloads are 493002. The API safetensors block reports BF16: 3085938688 and total: 3085938688. The model card states this is the base 3B Qwen2.5 model with 3.09B parameters, 2.77B non-embedding parameters, 36 layers, 16 Q heads, 2 KV heads, tied word embeddings, and 32768-token context." }, { "label": "Qwen2.5 3B config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/3aab1f1954e9cc14eb9509a215f9e5ca08227a9b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, model_type qwen2, bfloat16, tie_word_embeddings true, hidden_size 2048, intermediate_size 11008, 36 layers, 16 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, use_sliding_window false, max_window_layers 36, vocab_size 151936, use_cache true, and rope_theta 1000000." }, { "label": "Qwen2.5 3B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/3aab1f1954e9cc14eb9509a215f9e5ca08227a9b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 6171877376 bytes across two shards. Range-read shard headers found 434 BF16 tensors totaling 3085938688 parameters / 6.171877376 GB, matching index total_size. Linked-object HEAD checks resolved both shards to the pinned commit with combined linked size 6.171926992 GB and 49616 bytes of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 parameters / 0.622329856 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 3B Instruct comparison profile", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/aa8e72537993ba99e69dfaafa59ed015b17504d1/config.json", "source_type": "config", "supports": [ "architecture" ], "notes": "The existing audited instruction-tuned profile records matching architecture, context, tied-embedding layout, KV geometry, and safetensors byte split. This base profile still uses the base repo's own pinned config and shard headers directly." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, generation config, safetensors index metadata, linked-object HEAD checks, direct safetensors shard header grouping, and comparison against the existing Qwen2.5 3B Instruct profile." }, "notes": "This is a self-contained dense BF16 profile for ordinary text-decode profile-backed bounds on the base Qwen2.5 3B repo." }, { "id": "qwen--qwen2-5-72b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-72B-Instruct-AWQ", "title": "Qwen2.5 72B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2.5 72B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-72B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, model card, served AWQ config, and current base config comparison", "config_compatible": false, "notes": "The AWQ repo records Qwen/Qwen2.5-72B-Instruct as its quantized base model. Manual comparison with the current base config found matching architecture, layer count, attention geometry, context, and embedding fields, but the served AWQ config records intermediate_size 29696 while the current base config records 29568. This profile therefore uses the AWQ artifact config and direct tensor headers as authoritative." }, "architecture": { "canonical_architecture_id": "qwen2-5-72b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 72.957861888, "swept_params_b": 71.7121536, "auxiliary_resident_params_b": 1.245708288, "resident_weight_gb": 41.595518976, "swept_weight_gb": 39.1041024, "auxiliary_resident_weight_gb": 2.491416576, "resident_parameter_scope": "logical Qwen2.5 72B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, F16 biases, and unquantized F16 embeddings/head tensors. Logical parameter counts use the Hugging Face safetensors API accounting and header grouping; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 8192 divided by 64 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes. The model card notes optional YaRN extension to 131072 tokens, but the shipped config max_position_embeddings is 32768, so the default profile records 32768." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.570130728883676, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar. weight_bytes_per_param records resident stored bytes divided by reconstructed resident logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 72B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-72B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 698703eae6604af048a3d2f509995dc302088217, the API records a public Qwen-licensed text-generation repo with base_model Qwen/Qwen2.5-72B-Instruct, text-generation-inference, endpoints_compatible, 4-bit, AWQ, and region:us tags. Current downloads are 1056283. The API safetensors block reports logical parameters split across I32: 70464307200 and F16: 2493554688, total 72957861888." }, { "label": "Qwen2.5 72B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-AWQ/raw/698703eae6604af048a3d2f509995dc302088217/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen2ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 80 layers, 8 KV heads, hidden_size 8192, 64 attention heads, intermediate_size 29696, 32768 max position embeddings, tie_word_embeddings false, sliding_window 131072, max_window_layers 70, and use_sliding_window false." }, { "label": "Qwen2.5 72B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching checked architecture fields except intermediate_size: the current base config records 29568, while the AWQ artifact config and qweight tensor shapes record 29696. The profile uses the AWQ artifact geometry for the audited bound." }, { "label": "Qwen2.5 72B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-AWQ/raw/698703eae6604af048a3d2f509995dc302088217/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The index records total_size 41595518976 bytes across 11 shards. Range-read safetensors headers found 2083 tensors totaling 41.595518976 GB: 35.5074048 GB I32 tensors and 6.088114176 GB F16 tensors. Stored suffix bytes are qweight 35.2321536 GB, F16 weight 4.985470976 GB, F16 scales 1.1010048 GB, qzeros 0.2752512 GB, and F16 bias 0.0016384 GB. model.embed_tokens.weight is F16 with shape [152064, 8192] and contributes 1.245708288B logical parameters / 2.491416576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 39.1041024 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, served AWQ config, current base Instruct config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, and unquantized embedding/head tensors." }, { "id": "qwen--qwen2-5-72b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-72B-Instruct", "title": "Qwen2.5 72B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 72B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-72B", "relation": "finetune", "source": "Hugging Face model metadata, model card, and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings while the base config records 131072. The served Instruct config also differs in rms_norm_eps and eos_token_id, so this profile uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-72b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 72.706203648, "swept_params_b": 71.46049536, "auxiliary_resident_params_b": 1.245708288, "resident_weight_gb": 145.412407296, "swept_weight_gb": 142.92099072, "auxiliary_resident_weight_gb": 2.491416576, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 72706203648 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 8192 divided by 64 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes long-context support up to 128K, but the current served config sets max_position_embeddings to 32768 and the card states the current config is set for 32768 tokens, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 72B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-72B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit 495f39366efef23836d0cfae4fbe635880d2be31, the API records a public non-gated text-generation repo with base_model Qwen/Qwen2.5-72B, license other / qwen, text-generation-inference, endpoints_compatible, deploy:azure, region:us, and safetensors parameters BF16: 72706203648. Current downloads were 627874 when audited." }, { "label": "Qwen2.5 72B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 80 layers, hidden_size 8192, intermediate_size 29568, 64 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, rms_norm_eps 1e-6, and 32768 max position embeddings." }, { "label": "Qwen2.5 72B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct", "source_type": "model_card", "supports": [ "max_context_tokens", "license", "base_model_proof" ], "notes": "The card frontmatter records license other, license_name qwen, and base_model Qwen/Qwen2.5-72B. The card describes optional long-context deployment but says the current config.json is set for context length up to 32768 tokens." }, { "label": "Qwen2.5 72B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 145412407296 bytes across 37 shards. Range-read shard headers found 963 BF16 tensors totaling 72706203648 parameters / 145.412407296 GB, matching the index total. model.embed_tokens.weight has shape [152064, 8192] and contributes 1245708288 parameters / 2.491416576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 71460495360 parameters / 142.92099072 GB." }, { "label": "Qwen2.5 72B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B/raw/efba10c8e54e91e0d9570ab5f7b51a958474d4cb/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different max_position_embeddings, rms_norm_eps, and eos_token_id values, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, model card, base config comparison, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-7b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-7B-Instruct-AWQ", "title": "Qwen2.5 7B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2.5 7B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-7B-Instruct as its quantized base model. Manual comparison found matching Qwen2ForCausalLM tensor geometry and context settings; the AWQ artifact adds quantization_config and changes torch_dtype from bfloat16 to float16." }, "architecture": { "canonical_architecture_id": "qwen2-5-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 5.570747392, "swept_weight_gb": 4.48075264, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "logical Qwen2.5 7B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, F16 scales, F16 biases, and unquantized F16 embeddings/head tensors. Logical parameter counts match the BF16 base model; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and no modules_to_not_convert list. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 7B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-7B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "At commit b25037543e9394b818fdfca67ab2a00ecc7dd641, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-7B-Instruct and tags 4-bit and awq. The API safetensors block reports logical parameters split across I32: 6525288448 and F16: 1090328064, total 7615616512." }, { "label": "Qwen2.5 7B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ/raw/b25037543e9394b818fdfca67ab2a00ecc7dd641/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen2ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, 28 layers, 4 KV heads, hidden_size 3584, 28 attention heads, 32768 max position embeddings, tie_word_embeddings false, and use_sliding_window false." }, { "label": "Qwen2.5 7B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/a09a35458c702b33eeacc393d103063234e8bc28/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the base Instruct config after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version." }, { "label": "Qwen2.5 7B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ/raw/b25037543e9394b818fdfca67ab2a00ecc7dd641/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The index records total_size 5570747392 bytes across two shards. Range-read safetensors headers found 731 tensors totaling 5.570747392 GB: 3.288133632 GB I32 tensors and 2.28261376 GB F16 tensors. model.embed_tokens.weight is F16 with shape [152064, 3584] and contributes 544997376 logical parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 4.48075264 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served AWQ config, base Instruct config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, and unquantized embedding/head tensors." }, { "id": "qwen--qwen2-5-7b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-7B-Instruct-GGUF", "title": "Qwen2.5 7B Instruct GGUF Q2_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected Q2_K GGUF artifact of Qwen2.5 7B Instruct.", "model_family": "qwen2.5-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B-Instruct", "relation": "quantized", "source": "Hugging Face API metadata, audited BF16 Instruct profile, linked-object HEAD checks, and selected Q2_K GGUF header metadata", "config_compatible": true, "notes": "The official GGUF repo records Qwen/Qwen2.5-7B-Instruct as its quantized base. The selected Q2_K GGUF header records the same Qwen2ForCausalLM tensor geometry as the audited BF16 Instruct profile: 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128 key/value head dimension, and separate input/output embeddings. The BF16 Instruct repo config records 32768 max positions, while this selected GGUF header records 131072 context; this profile uses the selected serving artifact header for the GGUF row." }, "architecture": { "canonical_architecture_id": "qwen2-5-7b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 3.01594, "swept_weight_gb": 2.831159296, "auxiliary_resident_weight_gb": 0.184780704, "resident_parameter_scope": "selected qwen2.5-7b-instruct-q2_k.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.27 tensors from the selected Q2_K GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for each ordinary text decode token", "notes": "The audited BF16 Instruct profile stores separate model.embed_tokens.weight and lm_head.weight tensors. The selected Q2_K GGUF likewise stores token_embd.weight and output.weight separately. This profile charges output.weight as swept decode traffic and treats token_embd.weight as resident-only input embedding storage. The selected linked file is 3.015940000 GB. Header tensor spans total 3.009986560 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005953440 GB. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 2-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The audited BF16 Instruct config records use_sliding_window false. The selected GGUF header records 28 Qwen2 decoder layers, 4 KV heads, 128-dimensional key/value heads, and qwen2.context_length 131072. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected Q2_K GGUF artifact. FP16 and other quantized siblings in the repo have different resident and swept bytes and need separate selected-artifact profiles if they are exposed." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.3960204660053135, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q2-k-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, tokenizer processing, kernels, dequantization, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The live HF API gguf.totalFileSize exactly matches qwen2.5-7b-instruct-q2_k.gguf, so this profile targets the API-selected Q2_K artifact. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Qwen2.5 7B Instruct GGUF HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-7B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit bb5d59e06d9551d752d08b292a50eb208b07ab1f, the live API records an official public non-gated Apache-2.0 GGUF text-generation repo with base_model Qwen/Qwen2.5-7B-Instruct, base_model:quantized metadata, region:us, 99520 live downloads, GGUF architecture qwen2, context_length 131072, gguf.total 7615616512, and gguf.totalFileSize 3015940000. The API totalFileSize exactly matches qwen2.5-7b-instruct-q2_k.gguf, so this profile targets that artifact. The catalog retains the older qualifying scrape count 107559 until the over-100k working set is regenerated." }, { "label": "Qwen2.5 7B Instruct GGUF model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/raw/bb5d59e06d9551d752d08b292a50eb208b07ab1f/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "serving" ], "notes": "The pinned official repo card and metadata identify the package as a GGUF quantization of Qwen/Qwen2.5-7B-Instruct with Apache-2.0 licensing and llama.cpp/GGUF-style serving artifacts." }, { "label": "Qwen2.5 7B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "embedding_layout" ], "notes": "The existing audited BF16 profile records Qwen2ForCausalLM, 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128 head dimension, use_sliding_window false, BF16 safetensors, and separate input/output embeddings. It uses the BF16 repo config context of 32768, while this profile uses the selected GGUF header context of 131072." }, { "label": "Qwen2.5 7B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/tree/bb5d59e06d9551d752d08b292a50eb208b07ab1f", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found fp16 split parts totaling 15.237853536 GB, q2_k 3.015940000 GB, q3_k_m 3.808391072 GB, q4_0 split parts totaling 4.431390848 GB, q4_k_m split parts totaling 4.683073632 GB, q5_0 split parts totaling 5.315176576 GB, q5_k_m split parts totaling 5.444831360 GB, q6_k split parts totaling 6.254198880 GB, and q8_0 split parts totaling 8.098525408 GB. The q2_k linked-object size exactly matches API gguf.totalFileSize." }, { "label": "Qwen2.5 7B Instruct Q2_K GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/resolve/bb5d59e06d9551d752d08b292a50eb208b07ab1f/qwen2.5-7b-instruct-q2_k.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 26 metadata entries and 339 tensors. The linked file is 3.015940000 GB. Tensor spans sum to 3.009986560 GB across 7.615616512B logical elements: output.weight 0.447068160 GB, output_norm.weight 0.000014336 GB, token_embd.weight 0.178827264 GB, and blk.0-27 tensors 2.384076800 GB. Metadata/tokenizer/header/file overhead accounts for 0.005953440 GB. Tensor spans split into Q2_K 1.561276416 GB, Q3_K 0.971407360 GB, Q4_K 0.028901376 GB, Q6_K 0.447068160 GB, and F32 0.001333248 GB. The header records general.architecture qwen2, qwen2.block_count 28, context_length 131072, embedding_length 3584, feed_forward_length 18944, attention.head_count 28, attention.head_count_kv 4, rope.freq_base 1000000, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, existing BF16 base profile, linked GGUF file sizes, and a direct selected-Q2_K GGUF header/tensor-index range read." }, "notes": "Use this profile for the official Qwen Q2_K GGUF text artifact in ordinary text-decode bounds. Do not infer FP16 or other quantized sibling footprints from this profile." }, { "id": "qwen--qwen2-5-7b-instruct", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-7B-Instruct", "title": "Qwen2.5 7B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 7B Instruct repo.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings while the base config records 131072. This profile uses the Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "notes": "Range-read safetensors headers record 7615616512 BF16 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 7B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen2.5-7B as the base model and Apache-2.0 as the license." }, { "label": "Qwen2.5 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen2ForCausalLM, bfloat16, 28 layers, 4 KV heads, hidden_size 3584, 28 attention heads, 32768 max position embeddings, and use_sliding_window false." }, { "label": "Qwen2.5 7B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-7B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format" ], "notes": "At commit a09a35458c702b33eeacc393d103063234e8bc28, the API safetensors block records BF16: 7615616512 and total: 7615616512, which this profile stores as 7.615616512B resident parameters." }, { "label": "Qwen2.5 7B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/a09a35458c702b33eeacc393d103063234e8bc28/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "weight_format", "embedding_layout" ], "notes": "The index lists four safetensors shards. Range-read shard headers record 339 BF16 tensors totaling 7615616512 parameters and 15.231233024 GB, matching index total_size. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB." }, { "label": "Qwen2.5 7B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but a different max_position_embeddings value, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-7B", "title": "Qwen2.5 7B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2.5 7B base repo.", "model_family": "qwen2.5-dense", "architecture": { "canonical_architecture_id": "qwen2-5-7b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.231233024, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 7615616512 BF16 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM base profile using the served base repo config. The base config records 131072 max position embeddings, unlike the audited instruction-tuned repo config that records 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 7B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-7B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "architecture", "total_params_b" ], "notes": "At repo SHA d149729398750b98c0af14eb82c78cfe92750796, the API records a public Apache-2.0 text-generation Transformers safetensors repo with qwen2, conversational, eval-results, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 544225. The API safetensors block reports BF16: 7615616512 and total: 7615616512. The model card states this is the base 7B Qwen2.5 model with 7.61B parameters, 6.53B non-embedding parameters, 28 layers, 28 Q heads, 4 KV heads, and 131072-token context." }, { "label": "Qwen2.5 7B config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, model_type qwen2, bfloat16, tie_word_embeddings false, hidden_size 3584, intermediate_size 18944, 28 layers, 28 attention heads, 4 KV heads, 131072 max position embeddings, sliding_window 131072, use_sliding_window false, vocab_size 152064, and rope_theta 1000000." }, { "label": "Qwen2.5 7B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The index lists four safetensors shards and records total_size 15231233024 bytes. Range-read shard headers record 339 BF16 tensors totaling 7615616512 parameters and 15.231233024 GB, matching index total_size. Linked-object HEAD checks for all four shards resolved to commit d149729398750b98c0af14eb82c78cfe92750796, with combined linked file size 15.231271888 GB and 38864 bytes of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB." }, { "label": "Qwen2.5 7B Instruct comparison profile", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/a09a35458c702b33eeacc393d103063234e8bc28/config.json", "source_type": "config", "supports": [ "architecture" ], "notes": "The audited instruction-tuned profile has matching core tensor geometry and the same safetensors byte split, but its served config records 32768 max position embeddings. This base profile therefore uses the base repo's pinned 131072-token config directly." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, safetensors index metadata, linked-object HEAD checks, direct safetensors shard header grouping, and comparison against the existing Qwen2.5 7B Instruct profile." }, "notes": "This is a self-contained dense BF16 profile for ordinary text-decode profile-backed bounds on the base Qwen2.5 7B repo." }, { "id": "qwen--qwen2-5-coder-1-5b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-1.5B-Instruct", "title": "Qwen2.5 Coder 1.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 1.5B Instruct repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-1.5B", "relation": "finetune", "source": "Hugging Face model metadata and pinned base config comparison", "config_compatible": true, "notes": "The target repo records Qwen/Qwen2.5-Coder-1.5B as its base model. Manual comparison found matching audited tensor geometry, context fields, tied embedding setting, dtype, vocab size, and model type between the target config and the Coder base config; eos_token_id differs because this is the instruction-tuned checkpoint." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM Coder profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 1.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen2.5-Coder-1.5B as the base model and Apache-2.0 as the license. The card describes the 1.5B Coder Instruct checkpoint and lists the Qwen2.5 Coder architecture features." }, { "label": "Qwen2.5 Coder 1.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/raw/2e1fd397ee46e1388853d2af2c993145b0f1098a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The pinned config records Qwen2ForCausalLM, bfloat16, hidden_size 1536, intermediate_size 8960, 28 layers, 12 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, tie_word_embeddings true, and use_sliding_window false." }, { "label": "Qwen2.5 Coder 1.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-1.5B-Instruct", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 2e1fd397ee46e1388853d2af2c993145b0f1098a, the API records a public, non-gated Apache-2.0 text-generation repo with safetensors parameters BF16: 1543714304 and total: 1543714304. Current downloads were 805428 when audited." }, { "label": "Qwen2.5 Coder 1.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/resolve/2e1fd397ee46e1388853d2af2c993145b0f1098a/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 Coder 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/raw/df3ce67c0e24480f20468b6ef2894622d69eb73b/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching audited geometry and serving fields between the target Instruct config and the Coder base config, except for the instruction-tuned eos token." }, { "label": "Qwen2.5 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison against the broader Qwen2.5 1.5B lineage found the same core tensor geometry, with longer base context fields in Qwen/Qwen2.5-1.5B. This profile uses the served Coder Instruct repo config directly." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served config, pinned Coder base config comparison, Qwen2.5 lineage comparison, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-coder-1-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-1.5B", "title": "Qwen2.5 Coder 1.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 1.5B base repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B", "relation": "finetune", "source": "Hugging Face model metadata, Qwen2.5 base config comparison, and Coder Instruct sibling comparison", "config_compatible": false, "notes": "The model card and API metadata identify Qwen/Qwen2.5-1.5B as the base model. Manual comparison found matching core tensor geometry, dtype, vocabulary, tokenizer IDs, and tied-embedding fields with the Qwen2.5 1.5B base config, but the Coder repo records 32768 max position embeddings and sliding_window while the current Qwen2.5 base config records 131072. Manual comparison against the audited Coder Instruct sibling found matching tensor and context geometry; only eos_token_id differs." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Coder base repo config directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 1.5B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-1.5B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit df3ce67c0e24480f20468b6ef2894622d69eb73b, the API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen2, code, qwen-coder, codeqwen, text-generation-inference, endpoints_compatible, deploy:azure, region:us, and base_model Qwen/Qwen2.5-1.5B metadata; 189380 downloads; and safetensors parameters BF16: 1543714304, total: 1543714304." }, { "label": "Qwen2.5 Coder 1.5B config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/raw/df3ce67c0e24480f20468b6ef2894622d69eb73b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, tie_word_embeddings true, bfloat16, hidden_size 1536, intermediate_size 8960, 28 layers, 12 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, use_sliding_window false, max_window_layers 28, vocab_size 151936, eos_token_id 151643, and rope_theta 1000000." }, { "label": "Qwen2.5 Coder 1.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/resolve/df3ce67c0e24480f20468b6ef2894622d69eb73b/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 38528-byte header with 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.embed_tokens.weight has shape [151936, 1536] and contributes 233373696 parameters / 0.466747392 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. Header grouping found MLP tensors totaling 2.312110080 GB, self-attention tensors 0.308396032 GB, layer norms 0.000172032 GB, and final norm 0.000003072 GB." }, { "label": "Qwen2.5 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry, dtype, vocabulary, tokenizer IDs, and tied-embedding fields between Qwen/Qwen2.5-Coder-1.5B and Qwen/Qwen2.5-1.5B. The current upstream base differs in context fields: max_position_embeddings and sliding_window are 131072 in Qwen/Qwen2.5-1.5B and 32768 in the Coder base." }, { "label": "Qwen2.5 Coder 1.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/raw/2e1fd397ee46e1388853d2af2c993145b0f1098a/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family" ], "notes": "Manual comparison against the already audited Coder Instruct sibling found matching tensor geometry, dtype, vocabulary, tied embedding setting, and context fields. The only checked config difference is eos_token_id: 151643 for the base Coder repo and 151645 for the Instruct repo." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served Coder config, Qwen2.5 base config comparison, audited Coder Instruct config comparison, model card metadata, and a direct safetensors header range read." }, "notes": "This is a self-contained dense BF16 profile for the Qwen2.5 Coder 1.5B base repo, separate from the existing Coder Instruct profile." }, { "id": "qwen--qwen2-5-coder-14b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-14B-Instruct-AWQ", "title": "Qwen2.5 Coder 14B Instruct AWQ", "summary": "Audited memory-side bounds profile for the Qwen2.5 Coder 14B Instruct AWQ 4-bit checkpoint.", "model_family": "qwen2.5-coder-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-14B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served AWQ config, and audited BF16 base profile comparison", "config_compatible": true, "notes": "The AWQ repo card records Qwen/Qwen2.5-Coder-14B-Instruct as the base model. Manual comparison found matching Qwen2ForCausalLM geometry, context fields, sliding-window settings, vocabulary size, and tied-embedding setting between the AWQ and audited BF16 base configs. The only audited config-field difference is torch_dtype: float16 for AWQ and bfloat16 for the base." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 9.980028928, "swept_weight_gb": 8.422893568, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros, scales, and biases are storage/runtime overhead rather than logical model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 48 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes 128K long-context use through a YaRN config change, which is outside this default profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 Coder 14B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit eb3172f06a6d6b3a15f08947b0668d782e4d2d2c, the API reports a public/non-gated text-generation Transformers repo with base_model Qwen/Qwen2.5-Coder-14B-Instruct, Apache-2.0 license, tags 4-bit and awq, safetensors logical parameters I32: 13212057600, F16: 1557976064, total: 14770033664, and current downloads 1077947." }, { "label": "Qwen2.5 Coder 14B Instruct AWQ served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ/raw/eb3172f06a6d6b3a15f08947b0668d782e4d2d2c/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 13824, 48 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Qwen2.5 Coder 14B Instruct AWQ model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ/blob/eb3172f06a6d6b3a15f08947b0668d782e4d2d2c/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the AWQ-quantized 4-bit instruction-tuned 14B Qwen2.5-Coder model, records 48 layers and GQA with 40 Q heads and 8 KV heads, and notes that the current config is set for 32768 tokens while longer context requires adding YaRN rope scaling." }, { "label": "Qwen2.5 Coder 14B Instruct BF16 base config/profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/aedcc2d42b622764e023cf882b6652e646b95671/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout" ], "notes": "Manual comparison against the audited BF16 base profile found matching tensor geometry, context, sliding-window, vocabulary, and tied-embedding fields. The base profile records 14.770033664B logical parameters, 13.991465984B ordinary swept logical parameters, and a separate 0.778567680B input embedding tensor." }, { "label": "Qwen2.5 Coder 14B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ/raw/eb3172f06a6d6b3a15f08947b0668d782e4d2d2c/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 9980028928 bytes across three shards, matching direct range-read shard header spans. Headers contain 1251 tensors totaling 9.980028928 GB: 6.657638400 GB I32 tensors and 3.322390528 GB F16 tensors. Stored suffix totals are qweight 6.606028800 GB, qzeros 0.051609600 GB, scales 0.206438400 GB, F16 bias tensors 0.000688128 GB, and F16 weight tensors 3.115264000 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the served AWQ config, model card, audited BF16 14B base profile comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by using exact AWQ stored bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "qwen--qwen2-5-coder-14b-instruct", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-14B-Instruct", "title": "Qwen2.5 Coder 14B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 14B Instruct repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-14B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Manual comparison found matching tensor geometry, context fields, sliding-window settings, and dtype between the Instruct and base configs; eos_token_id differs, so this profile uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-14b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 29.540067328, "swept_weight_gb": 27.982931968, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 579 BF16 tensors totaling 14770033664 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes 128K long-context support via YaRN, but the current config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 14B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "architecture" ], "notes": "The card metadata identifies Qwen/Qwen2.5-Coder-14B as the base model and Apache-2.0 as the license. The card rounds the model as 14.7B parameters, 13.1B non-embedding parameters, 48 layers, 40 Q heads, 8 KV heads, and 128K long-context support via YaRN." }, { "label": "Qwen2.5 Coder 14B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/aedcc2d42b622764e023cf882b6652e646b95671/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 48 layers, hidden size 5120, 40 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, and 32768 max position embeddings." }, { "label": "Qwen2.5 Coder 14B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-14B-Instruct", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format", "base_model_proof" ], "notes": "The HF CLI model info response records repo SHA aedcc2d42b622764e023cf882b6652e646b95671, lastModified 2025-01-12T02:02:59Z, pipeline text-generation, base-model tags for Qwen/Qwen2.5-Coder-14B, license:apache-2.0, and safetensors parameters BF16: 14770033664." }, { "label": "Qwen2.5 Coder 14B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/aedcc2d42b622764e023cf882b6652e646b95671/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_size 29540067328 bytes across six shards. Range-read shard headers found 579 BF16 tensors totaling 14770033664 parameters / 29.540067328 GB, matching the index total. model.embed_tokens.weight has shape [152064, 5120] and contributes 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 27.982931968 GB." }, { "label": "Qwen2.5 Coder 14B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant tensor geometry, context, sliding-window, and dtype fields match the Instruct repo config; only eos_token_id differs." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from served config, base config comparison, model card/API metadata, safetensors index, range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-coder-32b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-32B-Instruct-AWQ", "title": "Qwen2.5 Coder 32B Instruct AWQ", "summary": "Audited memory-side bounds profile for the Qwen2.5 Coder 32B Instruct AWQ 4-bit checkpoint.", "model_family": "qwen2.5-coder-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata and served AWQ config", "config_compatible": true, "notes": "The AWQ repo card records Qwen/Qwen2.5-Coder-32B-Instruct as the base model. The served AWQ config records the same Qwen2ForCausalLM geometry as the audited BF16 base profile, with AWQ GEMM 4-bit quantization added." }, "architecture": { "canonical_architecture_id": "qwen2.5-coder-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 19.328804864, "swept_weight_gb": 17.771669504, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as their unpacked 4-bit logical parameters, while qzeros, scales, and biases are storage/runtime overhead rather than logical model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes 128K long-context use through a YaRN config change, which is outside this default profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 Coder 32B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 1ed0a6145da0ce550c628e8e8b678f51e695995d, the API reports a text-generation Transformers repo with base_model Qwen/Qwen2.5-Coder-32B-Instruct, tags 4-bit and awq, safetensors logical parameters I32: 31205621760, F16: 1558254592, total: 32763876352, and current downloads 1707036." }, { "label": "Qwen2.5 Coder 32B Instruct AWQ served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ/raw/1ed0a6145da0ce550c628e8e8b678f51e695995d/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Qwen2.5 Coder 32B Instruct AWQ model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ/blob/1ed0a6145da0ce550c628e8e8b678f51e695995d/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the AWQ-quantized 4-bit instruction-tuned 32B Qwen2.5-Coder model, records 64 layers and GQA with 40 Q heads and 8 KV heads, and notes that the default config is set for 32768 tokens while longer context requires adding YaRN rope scaling." }, { "label": "Qwen2.5 Coder 32B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ/raw/1ed0a6145da0ce550c628e8e8b678f51e695995d/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 19328804864 bytes across five shards, matching direct range-read shard header spans. Headers contain 1667 tensors totaling 19.328804864 GB: 15.72470784 GB I32 tensors and 3.604097024 GB F16 tensors. Stored suffix totals are qweight 15.60281088 GB, qzeros 0.12189696 GB, scales 0.48758784 GB, F16 bias tensors 0.000917504 GB, and F16 weight tensors 3.11559168 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.55713536 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the served AWQ config, model card, base Qwen2.5-Coder 32B Instruct profile comparison, safetensors index, and direct shard header byte grouping." }, "notes": "Audited from HF API metadata, served config, model card, local base-profile comparison, safetensors index metadata, and direct safetensors shard header range reads." }, { "id": "qwen--qwen2-5-coder-32b-instruct-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4", "title": "Qwen2.5 Coder 32B Instruct GPTQ Int4", "summary": "Audited memory-side bounds profile for the official Qwen2.5 Coder 32B Instruct GPTQ 4-bit checkpoint.", "model_family": "qwen2.5-coder-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card metadata, served GPTQ config, audited BF16 Coder base profile, and safetensors shard headers", "config_compatible": true, "notes": "The GPTQ repo card and API metadata record Qwen/Qwen2.5-Coder-32B-Instruct as the quantized base model. Manual comparison against the audited BF16 Coder Instruct config found matching checked architecture fields; the GPTQ repo changes torch_dtype to float16 and adds GPTQ 4-bit quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2.5-coder-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 19.343747072, "swept_weight_gb": 17.786611712, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "logical GPTQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the GPTQ package mixes packed I32 qweight/qzeros/g_idx tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the audited BF16 base profile and Hugging Face API total; exact resident/swept byte fields drive production bounds." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "The served config records 32768 max_position_embeddings. The model card describes full 131072-token context via YaRN configuration changes, which are outside this default artifact profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.590398610475137, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-gptq-exllama-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ packed weights, qzeros, g_idx tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. GPTQ dequantization, ExLlama kernels, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by HF API logical parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen2.5 Coder 32B Instruct GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 5fece1e232475f4da75cf0e4283606a3186fee8a, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen2.5-Coder-32B-Instruct, Apache-2.0 license, tags 4-bit and gptq, region:us, safetensors logical parameters I32: 31205621760 and F16: 1558254592, total: 32763876352, and current downloads 148914." }, { "label": "Qwen2.5 Coder 32B Instruct GPTQ Int4 served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/raw/5fece1e232475f4da75cf0e4283606a3186fee8a/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 5120, intermediate_size 27648, 64 layers, 40 attention heads, 8 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, and GPTQ 4-bit quantization with group_size 128, sym true, desc_act false, true_sequential true, and use_exllama true." }, { "label": "Qwen2.5 Coder 32B Instruct GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/blob/5fece1e232475f4da75cf0e4283606a3186fee8a/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 4-bit instruction-tuned 32B Qwen2.5-Coder model, records 64 layers, GQA with 40 Q heads and 8 KV heads, GPTQ 4-bit quantization, and notes that inputs beyond 32768 tokens require adding YaRN rope scaling." }, { "label": "Qwen2.5 Coder 32B Instruct audited BF16 profile", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "total_params_b", "kv_adapter", "embedding_layout" ], "notes": "The audited BF16 Coder Instruct profile records Qwen2ForCausalLM, BF16, 64 layers, hidden size 5120, intermediate size 27648, 40 attention heads, 8 KV heads, 128-dimensional key/value heads, use_sliding_window false, 32768-token context, 32763876352 logical parameters, and separate model.embed_tokens.weight plus lm_head.weight tensors." }, { "label": "Qwen2.5 Coder 32B Instruct GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/resolve/5fece1e232475f4da75cf0e4283606a3186fee8a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 19343747072 bytes across five shards, matching direct range-read shard header spans. Headers contain 2115 tensors totaling 19.343747072 GB: 15.739650048 GB I32 tensors and 3.604097024 GB F16 tensors. Stored suffix totals are qweight 15.602810880 GB, qzeros 0.121896960 GB, g_idx 0.014942208 GB, scales 0.487587840 GB, F16 bias tensors 0.000917504 GB, and F16 weight tensors 3.115591680 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 5120] and contribute 1.557135360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served GPTQ config, model card, audited BF16 Coder base-profile comparison, safetensors index metadata, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen2.5 Coder 32B Instruct GPTQ Int4 artifact. Do not substitute the AWQ or non-Coder Qwen2.5 32B GPTQ profiles; the geometry is shared but the commit and base model evidence are distinct." }, { "id": "qwen--qwen2-5-coder-32b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-32B-Instruct", "title": "Qwen2.5 Coder 32B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 32B Instruct repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-32B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Manual comparison found matching tensor geometry, context fields, sliding-window settings, tied-embedding setting, dtype, and vocabulary size between the Instruct and base configs. eos_token_id differs, so this profile uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 65.527752704, "swept_weight_gb": 63.970617344, "auxiliary_resident_weight_gb": 1.55713536, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 32763876352 BF16 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes 128K long-context support via YaRN, but the current config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 32B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "architecture", "total_params_b", "weight_format" ], "notes": "At commit 381fc969f78efac66bc87ff7ddeadb7e73c218a7, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-Coder-32B and safetensors parameters BF16: 32763876352. The model card identifies this as the instruction-tuned 32B Qwen2.5-Coder model, with 32.5B parameters and 128K long-context support via optional YaRN; the current config remains set to 32768 context." }, { "label": "Qwen2.5 Coder 32B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct/raw/381fc969f78efac66bc87ff7ddeadb7e73c218a7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 64 layers, hidden size 5120, 40 attention heads, 8 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, and 32768 max position embeddings." }, { "label": "Qwen2.5 Coder 32B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B/raw/2e12b5f7bc878d424d222e224ed40aee564ec45f/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, dtype, vocab size, sliding_window, use_sliding_window, max_position_embeddings, and tied-embedding setting. The base config differs only on eos_token_id, so this profile uses the Instruct config directly." }, { "label": "Qwen2.5 Coder 32B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct/raw/381fc969f78efac66bc87ff7ddeadb7e73c218a7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 65527752704 bytes across 14 shards. Range-read shard headers found 771 BF16 tensors totaling 32763876352 parameters / 65.527752704 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 5120] and contributes 778567680 parameters / 1.55713536 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 31985308672 parameters / 63.970617344 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, base config comparison, safetensors index, direct range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-coder-3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-3B-Instruct", "title": "Qwen2.5 Coder 3B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 3B Instruct repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-3B", "relation": "finetune", "source": "Hugging Face model metadata, pinned target/base config comparison, and Qwen2.5 3B lineage comparison", "config_compatible": true, "notes": "The target repo records Qwen/Qwen2.5-Coder-3B as its base model. Manual comparison found matching tensor geometry, context fields, tied embedding setting, dtype, vocabulary size, and model type between the target config and the Coder base config. The broader Qwen/Qwen2.5-3B base config also has the same profile-relevant geometry." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 3.085938688, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.311164928, "non_embedding_params_b": 2.77477376, "notes": "Range-read safetensors headers record 3085938688 BF16 stored parameters. model.embed_tokens.weight contributes 311164928 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Dense Qwen2ForCausalLM Coder profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the two safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 3B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "model_card_claims" ], "notes": "The model metadata identifies Qwen/Qwen2.5-Coder-3B as the base model and qwen-research as the license. The card describes the instruction-tuned 3B Coder checkpoint with 3.09B parameters, 2.77B non-embedding parameters, 36 layers, 16 Q heads, 2 KV heads, tied word embeddings, and 32768-token context." }, { "label": "Qwen2.5 Coder 3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/raw/488639f1ff808d1d3d0ba301aef8c11461451ec5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The pinned config records Qwen2ForCausalLM, bfloat16, hidden_size 2048, intermediate_size 11008, 36 layers, 16 attention heads, 2 KV heads, 32768 max position embeddings, sliding_window 32768, tie_word_embeddings true, and use_sliding_window false." }, { "label": "Qwen2.5 Coder 3B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-3B-Instruct", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 488639f1ff808d1d3d0ba301aef8c11461451ec5, the API records a public non-gated text-generation repo with transformers, safetensors, qwen2, qwen-coder, region:us, and qwen-research license metadata. Current downloads were 226000 when audited. The safetensors block reports BF16: 3085938688 and total: 3085938688." }, { "label": "Qwen2.5 Coder 3B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/raw/488639f1ff808d1d3d0ba301aef8c11461451ec5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "The index lists two safetensors shards and total_size 6171877376 bytes. Range-read shard headers found 434 BF16 tensors totaling 3085938688 parameters / 6.171877376 GB. Shard header lengths were 39400 and 10208 bytes. model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 parameters / 0.622329856 GB. The index has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen2.5 Coder 3B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching audited geometry and serving fields between the target Instruct config and the Coder base config." }, { "label": "Qwen2.5 3B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison against the broader Qwen2.5 3B lineage found the same profile-relevant tensor geometry and context settings. This profile uses the served Coder Instruct repo config directly." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served config, Coder base config comparison, Qwen2.5 3B lineage comparison, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-coder-7b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-7B-Instruct-AWQ", "title": "Qwen2.5 Coder 7B Instruct AWQ", "summary": "Audited memory-side bounds profile for the official Qwen2.5 Coder 7B Instruct AWQ 4-bit checkpoint.", "model_family": "qwen2.5-coder-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-7B-Instruct", "relation": "quantized", "source": "Hugging Face API metadata, model card, served AWQ config, audited BF16 base profile comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The API metadata and model card identify Qwen/Qwen2.5-Coder-7B-Instruct as the base model. Manual comparison against the audited BF16 base profile found matching Qwen2ForCausalLM geometry, context fields, sliding-window settings, vocabulary size, and untied embedding setting; the AWQ repo changes torch_dtype to float16 and adds AWQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 5.570747392, "swept_weight_gb": 4.48075264, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer qweight/qzeros/scales/bias tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales/biases, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros, scales, and biases are storage/runtime overhead rather than logical model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 28 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "Dense Qwen2ForCausalLM AWQ profile using the served quantized repo config. The model card describes 128K long-context use through YaRN config changes, but the current config sets max_position_embeddings to 32768." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored AWQ bytes from safetensors headers: packed I32 qweight/qzeros tensors plus F16 scales, biases, embeddings, output head, and norms. Dequantization, activation traffic, compute, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged as FP16." }, "evidence": [ { "label": "Qwen2.5 Coder 7B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 8e8ed243bbe6f9a5aff549a0924562fc719b2b8a, the API reports a public non-gated Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen2.5-Coder-7B-Instruct, tags 4-bit, awq, deploy:azure, and region:us, current downloads 258690, and safetensors logical parameters I32: 6525288448, F16: 1090328064, total: 7615616512." }, { "label": "Qwen2.5 Coder 7B Instruct AWQ model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the AWQ-quantized 4-bit instruction-tuned 7B Qwen2.5-Coder model, records 7.61B parameters, 6.53B non-embedding parameters, 28 layers, GQA with 28 Q heads and 4 KV heads, AWQ 4-bit quantization, and says the current config is set for 32768 tokens while 128K context requires adding YaRN rope scaling." }, { "label": "Qwen2.5 Coder 7B Instruct AWQ served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ/raw/8e8ed243bbe6f9a5aff549a0924562fc719b2b8a/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 3584, intermediate_size 18944, 28 layers, 28 attention heads, 4 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, vocab_size 152064, rope_theta 1000000, and AWQ quantization with bits 4, group_size 128, version gemm, and zero_point true." }, { "label": "Qwen2.5 Coder 7B Instruct BF16 base profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/c03e6d358207e414f1eca0bb1891e29f1db0e242/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout" ], "notes": "Manual comparison against the audited BF16 base profile found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, context, sliding-window, vocabulary, and tied-embedding fields. The base profile records 7.615616512B logical parameters, 7.070619136B ordinary swept logical parameters, and a separate 0.544997376B input embedding tensor." }, { "label": "Qwen2.5 Coder 7B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ/raw/8e8ed243bbe6f9a5aff549a0924562fc719b2b8a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 5570747392 bytes across two shards. Direct range reads found shard sizes 3996422976 and 1574406784 bytes, with 82368 bytes of safetensors header overhead. Range-read shard headers contain 731 tensors totaling 5.570747392 GB: I32 3.288133632 GB and F16 2.282613760 GB. Stored suffix bytes are qweight 3.262644224 GB, qzeros 0.025489408 GB, scales 0.101957632 GB, F16 weight tensors 2.180398080 GB, and F16 bias tensors 0.000258048 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 3584] and contribute 1.089994752 GB. Ordinary swept traffic is layers, model.norm.weight, and lm_head.weight, totaling 4.480752640 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned AWQ config, audited BF16 base profile comparison, safetensors index, linked-object size checks, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by using exact AWQ stored bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "qwen--qwen2-5-coder-7b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-7B-Instruct-GGUF", "title": "Qwen2.5 Coder 7B Instruct GGUF Q5_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-example Q5_K_M GGUF artifact of Qwen2.5 Coder 7B Instruct.", "model_family": "qwen2.5-coder-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-7B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, audited BF16 base profile, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The official GGUF repo records Qwen/Qwen2.5-Coder-7B-Instruct as its quantized base. The selected Q5_K_M GGUF header records the same Qwen2ForCausalLM tensor geometry as the audited BF16 Instruct profile: 28 layers, 28 attention heads, 4 KV heads, 128 key/value head dimension, separate input/output embeddings, and full-context attention." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-7b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 5.444831232, "swept_weight_gb": 5.064192, "auxiliary_resident_weight_gb": 0.380639232, "resident_parameter_scope": "selected qwen2.5-coder-7b-instruct-q5_k_m.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.27 tensors from the selected Q5_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "The audited BF16 base profile stores separate model.embed_tokens.weight and lm_head.weight tensors. The selected Q5_K_M GGUF likewise stores token_embd.weight and output.weight separately. This profile charges output.weight as swept decode traffic and treats token_embd.weight as resident-only input embedding storage. The selected linked file is 5.444831232 GB. Header tensor spans total 5.438877696 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.005953536 GB." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The audited BF16 base config records use_sliding_window false. The selected GGUF header records 28 Qwen2 decoder layers, 4 KV heads, and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected Q5_K_M GGUF artifact. The selected GGUF header records 131072 context, while the card text says GGUF supports full 32768 context and recommends non-GGUF models for 131072-token processing; this profile uses the selected artifact header as serving truth and records the mismatch explicitly." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7149560673676947, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q5-k-m-qwen2.5-coder-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The model card's download example selects qwen2.5-coder-7b-instruct-q5_k_m*.gguf, while the live HF API gguf.totalFileSize matches qwen2.5-coder-7b-instruct-fp16.gguf. This profile intentionally targets the card-example Q5_K_M artifact." }, "evidence": [ { "label": "Qwen2.5 Coder 7B Instruct GGUF HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "The live HF API response at commit 13fb94bfda8c8cf22497dc57b78f391a9acb426a records an official Qwen GGUF repo with base_model Qwen/Qwen2.5-Coder-7B-Instruct, Apache-2.0 license, text-generation pipeline, region:us, 155704 downloads, GGUF architecture qwen2, 131072 context length, gguf.total 7615616512, and gguf.totalFileSize 15237853184. The API totalFileSize matches qwen2.5-coder-7b-instruct-fp16.gguf, while this profile targets Q5_K_M because the card's CLI download example selects that file." }, { "label": "Qwen2.5 Coder 7B Instruct GGUF model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/raw/13fb94bfda8c8cf22497dc57b78f391a9acb426a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "serving", "context_mismatch" ], "notes": "The pinned card records Apache-2.0 licensing, text-generation pipeline, base_model Qwen/Qwen2.5-Coder-7B-Instruct, Qwen2ForCausalLM-style architecture, 7.61B parameters, 6.53B non-embedding parameters, 28 layers, 28 query heads, 4 KV heads, 32768-token GGUF context guidance, available GGUF quantizations, and a Hugging Face CLI example downloading qwen2.5-coder-7b-instruct-q5_k_m*.gguf." }, { "label": "Qwen2.5 Coder 7B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "embedding_layout" ], "notes": "The existing audited BF16 profile records Qwen2ForCausalLM, 28 layers, 4 KV heads, 128 head dimension, use_sliding_window false, BF16 safetensors, and separate input/output embeddings." }, { "label": "Qwen2.5 Coder 7B Instruct GGUF linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/tree/13fb94bfda8c8cf22497dc57b78f391a9acb426a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found fp16 15.237853184 GB, q2_k 3.015940032 GB, q3_k_m 3.808391104 GB, q4_0 4.431390720 GB, q4_k_m 4.683073536 GB, q5_k_m 5.444831232 GB, q6_k 6.254198784 GB, and q8_0 8.098525184 GB. The split q5_k_m pair is 3.989841792 GB plus 1.454989568 GB, matching the merged Q5_K_M file size." }, { "label": "Qwen2.5 Coder 7B Instruct Q5_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/resolve/13fb94bfda8c8cf22497dc57b78f391a9acb426a/qwen2.5-coder-7b-instruct-q5_k_m.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 29 metadata entries and 339 tensors. The linked file is 5.444831232 GB. Tensor spans sum to 5.438877696 GB across 7.615616512B logical elements: output.weight 0.447068160 GB, output_norm.weight 0.000014336 GB, token_embd.weight 0.374685696 GB, and blk.0-27 tensors 4.617109504 GB. Metadata/tokenizer/header/file overhead accounts for 0.005953536 GB. Tensor spans split into Q5_K 4.189667328 GB, Q6_K 1.247877120 GB, and F32 0.001333248 GB. The header records qwen2.block_count 28, context_length 131072, embedding_length 3584, feed_forward_length 18944, attention.head_count 28, attention.head_count_kv 4, rope.freq_base 1000000, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, existing BF16 base profile, linked GGUF file sizes, and a direct selected-GGUF header/tensor-index range read." }, "notes": "Use this profile for the official Qwen Q5_K_M GGUF text artifact in ordinary text-decode bounds. Other GGUF quantizations in this repo have different resident and traffic bytes and require separate workload selection." }, { "id": "qwen--qwen2-5-coder-7b-instruct-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4", "title": "Qwen2.5 Coder 7B Instruct GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int4 Qwen2.5 Coder 7B Instruct repo.", "model_family": "qwen2.5-coder-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-7B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, served GPTQ config, audited BF16 base profile comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The API metadata and model card identify Qwen/Qwen2.5-Coder-7B-Instruct as the base model. Manual comparison against the audited BF16 base profile found matching Qwen2ForCausalLM geometry, context fields, sliding-window settings, vocabulary size, and untied embedding setting; the GPTQ repo changes torch_dtype to float16 and adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 5.575277568, "swept_weight_gb": 4.485282816, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "GPTQ logical serving parameters reconstructed from qweight tensors plus F16 model tensors in indexed safetensors shards", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer qweight/qzeros/g_idx/scales/bias tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full input embedding matrix for each generated token", "notes": "GPTQ qweight tensors are packed I32 values; logical parameter counts treat each qweight element as eight 4-bit values. qzeros, g_idx, and scales are storage/serving metadata and are charged in stored-byte traffic but excluded from logical parameter counts. F16 model weights and F16 attention biases are included as logical serving parameters." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records use_sliding_window false, so this profile charges full-context K and V streams for all 28 language layers. The config includes sliding_window 131072, but it is disabled by use_sliding_window false." }, "notes": "Dense Qwen2ForCausalLM GPTQ profile using the served quantized repo config. The model card describes 128K long-context use through YaRN config changes, but the current config sets max_position_embeddings to 32768." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.732084862626024, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-gptq-int4-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored GPTQ bytes from safetensors headers: packed I32 qweight/qzeros/g_idx tensors plus F16 scales, embeddings, output head, norms, and attention biases. Dequantization, activation traffic, compute, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 4-bit quantization with group_size 128, desc_act false, symmetric quantization, true_sequential true, and exllama v1 metadata. KV cache is charged as FP16." }, "evidence": [ { "label": "Qwen2.5 Coder 7B Instruct GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 3d0e91af138113d0b17d8011adf140d22423d13e, the API reports a public Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen2.5-Coder-7B-Instruct, tags 4-bit, gptq, deploy:azure, and region:us, current downloads 527068, and safetensors logical parameters I32: 6525288448, F16: 1090328064, total: 7615616512." }, { "label": "Qwen2.5 Coder 7B Instruct GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 4-bit instruction-tuned 7B Qwen2.5-Coder model, records 7.61B parameters, 6.53B non-embedding parameters, 28 layers, GQA with 28 Q heads and 4 KV heads, GPTQ 4-bit quantization, and says the current config is set for 32768 tokens while 128K context requires adding YaRN rope scaling." }, { "label": "Qwen2.5 Coder 7B Instruct GPTQ Int4 served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4/raw/3d0e91af138113d0b17d8011adf140d22423d13e/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen2ForCausalLM with hidden_size 3584, intermediate_size 18944, 28 layers, 28 attention heads, 4 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 131072, tie_word_embeddings false, torch_dtype float16, vocab_size 152064, rope_theta 1000000, and GPTQ quantization with bits 4, group_size 128, desc_act false, symmetric quantization, and true_sequential true." }, { "label": "Qwen2.5 Coder 7B Instruct BF16 base profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/c03e6d358207e414f1eca0bb1891e29f1db0e242/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout" ], "notes": "Manual comparison against the audited BF16 base profile found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, context, sliding-window, vocabulary, and tied-embedding fields. The base profile records 7.615616512B logical parameters, 7.070619136B ordinary swept logical parameters, and a separate 0.544997376B input embedding tensor." }, { "label": "Qwen2.5 Coder 7B Instruct GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4/raw/3d0e91af138113d0b17d8011adf140d22423d13e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index records total_size 5575277568 bytes across two shards. Direct HEAD checks found linked shard sizes 3999234216 and 1576146928 bytes, with 103576 bytes of safetensors header overhead. Range-read shard headers contain 927 tensors totaling 5.575277568 GB: I32 3.292663808 GB and F16 2.282613760 GB. Stored suffix bytes are qweight 3.262644224 GB, qzeros 0.025489408 GB, g_idx 0.004530176 GB, scales 0.101957632 GB, F16 weight tensors 2.180398080 GB, and F16 bias tensors 0.000258048 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 3584] and contribute 1.089994752 GB. Ordinary swept traffic is layers, model.norm.weight, and lm_head.weight, totaling 4.485282816 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned GPTQ config, audited BF16 base profile comparison, safetensors index, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by using exact GPTQ stored bytes and separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "qwen--qwen2-5-coder-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-7B-Instruct", "title": "Qwen2.5 Coder 7B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 7B Instruct repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Coder-7B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Manual comparison found matching tensor geometry, context fields, sliding-window settings, and dtype between the Instruct and base configs; eos_token_id differs, so this profile uses the served Instruct config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.231233024, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 339 BF16 tensors totaling 7615616512 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config. The model card describes 128K long-context support via YaRN, but the current config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 7B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "architecture", "total_params_b", "weight_format" ], "notes": "At commit c03e6d358207e414f1eca0bb1891e29f1db0e242, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-Coder-7B and safetensors parameters BF16: 7615616512. The model card identifies this as the instruction-tuned 7B Qwen2.5-Coder model, with 7.61B parameters, 28 layers, 28 Q heads, 4 KV heads, and 128K long-context support via optional YaRN." }, { "label": "Qwen2.5 Coder 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/c03e6d358207e414f1eca0bb1891e29f1db0e242/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 3584, 28 attention heads, 4 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, and 32768 max position embeddings." }, { "label": "Qwen2.5 Coder 7B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B/raw/0396a76181e127dfc13e5c5ec48a8cee09938b02/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, dtype, vocab size, sliding_window, use_sliding_window, max_position_embeddings, and tied-embedding setting. The base config differs only on eos_token_id, so this profile uses the Instruct config directly." }, { "label": "Qwen2.5 Coder 7B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/c03e6d358207e414f1eca0bb1891e29f1db0e242/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout" ], "notes": "The index records total_size 15231233024 bytes across four shards. Range-read shard headers found 339 BF16 tensors totaling 7615616512 parameters / 15.231233024 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, base config comparison, safetensors index, direct range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-5-coder-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Coder-7B", "title": "Qwen2.5 Coder 7B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Coder 7B base repo.", "model_family": "qwen2.5-coder-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-7B", "relation": "finetune", "source": "Hugging Face model card base_model metadata, served config, Qwen2.5-7B base config comparison, and direct safetensors header metadata", "config_compatible": false, "notes": "The repo metadata identifies Qwen/Qwen2.5-7B as the base model. Manual comparison found matching tensor geometry, dtype, sliding-window settings, vocabulary size, RoPE theta, and untied embedding layout, but the served Coder config sets max_position_embeddings to 32768 while the current Qwen2.5-7B base config sets it to 131072. This profile therefore uses the served Coder config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-coder-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.231233024, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 339 BF16 tensors totaling 7615616512 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served base Coder repo config. The model card describes 128K long-context support via YaRN, but the current config sets max_position_embeddings to 32768, so this v1 profile uses 32768." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Coder 7B API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "architecture", "total_params_b", "weight_format", "downloads", "commit_sha" ], "notes": "At commit 0396a76181e127dfc13e5c5ec48a8cee09938b02, the API records a public Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen2.5-7B, safetensors BF16 parameters 7615616512, current downloads 450017, endpoints_compatible, deploy:azure, and region:us tags." }, { "label": "Qwen2.5 Coder 7B model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B", "source_type": "model_card", "supports": [ "base_model_proof", "architecture", "total_params_b", "layers", "kv_heads", "context_notes" ], "notes": "The card identifies this repo as the 7B Qwen2.5-Coder base model and states 7.61B parameters, 6.53B non-embedding parameters, 28 layers, GQA with 28 Q heads and 4 KV heads, and long-context support up to 131072 tokens. The same card states the current config is set to 32768 tokens and requires adding YaRN rope_scaling for longer contexts." }, { "label": "Qwen2.5 Coder 7B config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B/raw/0396a76181e127dfc13e5c5ec48a8cee09938b02/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, vocab size 152064, rope_theta 1000000, and 32768 max position embeddings." }, { "label": "Qwen2.5 7B base config comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, dtype, vocabulary size, sliding_window, use_sliding_window, RoPE theta, and tied-embedding setting. The current Qwen2.5-7B base config sets max_position_embeddings to 131072 while the served Coder config sets 32768." }, { "label": "Qwen2.5 Coder 7B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B/raw/0396a76181e127dfc13e5c5ec48a8cee09938b02/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "dtype_split" ], "notes": "The index records total_size 15231233024 bytes across four shards. Direct HEAD checks found linked shard sizes 4877660776, 4932751008, 4330865200, and 1089994880 bytes, with 38840 bytes of safetensors header overhead. Range-read shard headers found 339 BF16 tensors totaling 7615616512 parameters / 15.231233024 GB, matching the index total and HF API safetensors total. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, Qwen2.5-7B base config comparison, safetensors index, direct HEAD checks, direct range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It supersedes the scraped metadata estimate by separating resident-only input embedding bytes from ordinary per-token swept decode traffic." }, { "id": "qwen--qwen2-5-math-1-5b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Math-1.5B-Instruct", "title": "Qwen2.5 Math 1.5B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Math 1.5B Instruct repo.", "model_family": "qwen2.5-math-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Math-1.5B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "The model metadata identifies Qwen/Qwen2.5-Math-1.5B as the base model. Manual comparison found matching architecture, tensor geometry, context, dtype, vocabulary, and tied-embedding fields between the base and Instruct configs." }, "architecture": { "canonical_architecture_id": "qwen2-5-math-1-5b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "model.safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A direct range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM math-instruction profile using the served Instruct repo config directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Math 1.5B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "serving" ], "notes": "The current Hugging Face metadata identifies the repo as an Apache-2.0 text-generation model derived from Qwen/Qwen2.5-Math-1.5B." }, { "label": "Qwen2.5 Math 1.5B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct/raw/aafeb0fc6f22cbf0eaeed126eff8be45b0360a35/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, bfloat16, hidden size 1536, intermediate size 8960, 28 layers, 12 attention heads, 2 KV heads, 4096 max position embeddings, sliding_window 4096, use_sliding_window false, tied embeddings, and vocab size 151936." }, { "label": "Qwen2.5 Math 1.5B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-1.5B-Instruct", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "The current API response records commit aafeb0fc6f22cbf0eaeed126eff8be45b0360a35, 484721 downloads when audited, public non-gated transformers text-generation metadata, region:us, and safetensors parameters BF16: 1543714304." }, { "label": "Qwen2.5 Math 1.5B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct/resolve/aafeb0fc6f22cbf0eaeed126eff8be45b0360a35/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A direct range read of the 38528-byte safetensors header found 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.layers total 2.620678144 GB, model.norm.weight totals 0.000003072 GB, model.embed_tokens.weight is [151936, 1536] and contributes 0.466747392 GB, and the checkpoint has no lm_head.weight." }, { "label": "Qwen2.5 Math 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/raw/4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited architecture, tensor-geometry, context, dtype, vocabulary, and tied-embedding fields." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned served config, model card/base-model metadata, base config comparison, and a direct safetensors header range read." }, "notes": "This profile supersedes the scraped metadata estimate with exact BF16 tensor payload and tied-embedding evidence." }, { "id": "qwen--qwen2-5-math-1-5b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Math-1.5B", "title": "Qwen2.5 Math 1.5B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2.5 Math 1.5B base repo.", "model_family": "qwen2.5-math-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B", "relation": "finetune", "source": "Hugging Face model metadata, served config, upstream Qwen2.5 base config comparison, and safetensors header review", "config_compatible": false, "notes": "The model card identifies Qwen/Qwen2.5-1.5B as the base model. Manual comparison found matching tensor geometry, dtype, vocabulary, and tied-embedding fields, but the Math repo sets max_position_embeddings and sliding_window to 4096 while Qwen/Qwen2.5-1.5B sets both to 131072. This profile therefore uses the served Math config directly." }, "architecture": { "canonical_architecture_id": "qwen2-5-math-1-5b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense", "total_params_b": 1.543714304, "parameter_scope": "model.safetensors_header_stored_bf16", "embedding_params_b": 0.233373696, "non_embedding_params_b": 1.310340608, "notes": "A direct range-read of model.safetensors records 1543714304 BF16 stored parameters. model.embed_tokens.weight contributes 233373696 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Dense Qwen2ForCausalLM math base profile using the served repo config directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 Math 1.5B model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "serving" ], "notes": "The current Hugging Face metadata identifies the repo as an Apache-2.0 text-generation math model derived from Qwen/Qwen2.5-1.5B." }, { "label": "Qwen2.5 Math 1.5B config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/raw/4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen2ForCausalLM, bfloat16, hidden size 1536, intermediate size 8960, 28 layers, 12 attention heads, 2 KV heads, 4096 max position embeddings, sliding_window 4096, use_sliding_window false, tied embeddings, and vocab size 151936." }, { "label": "Qwen2.5 Math 1.5B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-1.5B", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license" ], "notes": "The current API response records commit 4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2, 207753 downloads when audited, public non-gated transformers text-generation metadata, region:us, and safetensors parameters BF16: 1543714304." }, { "label": "Qwen2.5 Math 1.5B safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/resolve/4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A direct range read of the 38528-byte safetensors header found 338 BF16 tensors totaling 1543714304 parameters / 3.087428608 GB. model.layers total 2.620678144 GB, model.norm.weight totals 0.000003072 GB, model.embed_tokens.weight is [151936, 1536] and contributes 0.466747392 GB, and the checkpoint has no lm_head.weight." }, { "label": "Qwen2.5 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching tensor geometry, dtype, vocabulary, and tied-embedding fields between Qwen/Qwen2.5-1.5B and Qwen/Qwen2.5-Math-1.5B. The upstream base differs in context fields: max_position_embeddings and sliding_window are 131072 in Qwen/Qwen2.5-1.5B and 4096 in Qwen/Qwen2.5-Math-1.5B." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served config, model card/base-model metadata, upstream base config comparison, and a direct safetensors header range read." }, "notes": "This profile supersedes the scraped metadata estimate with exact BF16 tensor payload and tied-embedding evidence." }, { "id": "qwen--qwen2-5-math-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Math-7B-Instruct", "title": "Qwen2.5 Math 7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2.5 Math 7B Instruct safetensors checkpoint.", "model_family": "qwen2-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Math-7B", "relation": "finetune", "source": "Hugging Face model card/API metadata and served Qwen2 config", "config_compatible": true, "notes": "The model card/API metadata identify Qwen/Qwen2.5-Math-7B as the base model and this repo as the instruction-tuned derivative. The served config directly records the Qwen2ForCausalLM geometry used by this profile." }, "architecture": { "canonical_architecture_id": "qwen2.5-math-7b", "max_context_tokens": 4096, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.231233024, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 1.089994752, "resident_parameter_scope": "safetensors_header_bf16_payload", "swept_parameter_scope": "ordinary text decode charges model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each generated token because tie_word_embeddings is false and a separate lm_head.weight tensor is stored", "notes": "Range-read safetensors headers found 339 BF16 tensors totaling 15.231233024 GB. model.embed_tokens.weight and lm_head.weight are both present and each stores 1.089994752 GB; only lm_head.weight is charged as full output-projection traffic during ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false, 28 layers, 4 KV heads, 128 head dimension, BF16 dtype, and 4096 max positions. Although sliding_window and max_window_layers fields are present, the disabled flag means ordinary serving uses full-context KV for all 28 layers." }, "notes": "Qwen2ForCausalLM dense decoder profile with untied input/output embeddings." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-bf16-qwen2-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored BF16 safetensors payload bytes and BF16 full-context K/V cache traffic. Activation traffic, attention kernels, compute throughput, cache writes, and scheduler behavior are outside this memory-side bound.", "notes": "The API safetensors block records BF16 7615616512 parameters. The direct shard headers match 15.231233024 GB of BF16 payload bytes." }, "evidence": [ { "label": "Qwen2.5 Math 7B Instruct API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit ef9926d75ab1d54532f6a30dd5e760355eb9aa4d, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, qwen2, base_model Qwen/Qwen2.5-Math-7B, region:us, 148342 downloads, and safetensors BF16 total 7615616512." }, { "label": "Qwen2.5 Math 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct/raw/ef9926d75ab1d54532f6a30dd5e760355eb9aa4d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "embedding_layout" ], "notes": "The config records Qwen2ForCausalLM, model_type qwen2, bfloat16 dtype, 28 hidden layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 4096 max positions, use_sliding_window false, sliding_window 4096, max_window_layers 28, vocab size 152064, and tie_word_embeddings false." }, { "label": "Qwen2.5 Math 7B Instruct safetensors index and headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct/resolve/ef9926d75ab1d54532f6a30dd5e760355eb9aa4d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The safetensors index records total_size 15231233024 bytes across four shards. Linked object HEAD checks sum to 15.231271888 GB, with 0.000038864 GB of safetensors header/container overhead. Direct shard-header range reads found 339 BF16 tensors with payload bytes 15.231233024 GB. Swept ordinary text traffic is 14.141238272 GB: model.layers.* 13.051236352 GB, model.norm.weight 0.000007168 GB, and lm_head.weight 1.089994752 GB. Resident-only model.embed_tokens.weight is 1.089994752 GB." }, { "label": "Qwen2.5 Math 7B base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The API/card metadata for the instruct repo identify this as a finetune of Qwen/Qwen2.5-Math-7B. The served config and tensor headers in the instruct repo are the authoritative profile inputs." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, served config, safetensors index, linked-object HEAD checks, and direct safetensors shard header range reads." }, "notes": "This profile supersedes the generated dense estimate by separating resident input embedding bytes from swept ordinary text-decode traffic while preserving the Qwen2 full-context BF16 KV geometry." }, { "id": "qwen--qwen2-5-omni-3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Omni-3B", "title": "Qwen2.5 Omni 3B", "summary": "Audited memory-side bounds profile for ordinary text decode through the Qwen2.5-Omni 3B Thinker text decoder while charging the full Thinker/Talker/Token2Wav checkpoint as resident.", "model_family": "qwen2.5-omni-dense", "architecture": { "canonical_architecture_id": "qwen2.5-omni-3b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 5.537120672, "swept_params_b": 3.085938688, "auxiliary_resident_params_b": 2.451181984, "resident_weight_gb": 11.972343936, "swept_weight_gb": 6.171877376, "auxiliary_resident_weight_gb": 5.80046656, "resident_parameter_scope": "safetensors_header_stored_bf16_f32_full_thinker_talker_token2wav_package", "swept_parameter_scope": "thinker.model text decoder excluding thinker.model.embed_tokens.weight plus thinker.lm_head.weight", "auxiliary_scope": "thinker.model.embed_tokens.weight, thinker.audio_tower tensors, thinker.visual tensors, all talker tensors, and all token2wav tensors are resident for the full Omni checkpoint but not swept as full matrices for each generated text token", "notes": "The config records tie_word_embeddings false, and the headers store separate thinker.model.embed_tokens.weight and thinker.lm_head.weight tensors with identical shape. Ordinary text decode excludes the input embedding lookup and includes thinker.lm_head.weight, so the swept bytes equal the full thinker.model byte total. Talker and Token2Wav are retained in resident memory because the catalog row targets the full checkpoint; a talker-disabled deployment should be represented separately." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Thinker text config records use_sliding_window false, so ordinary text decode charges full-context K and V streams for all 36 Thinker text layers. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "The full Qwen2.5-Omni checkpoint contains a Thinker text/vision/audio model, a Talker model, and Token2Wav waveform modules. Bounds Engine v1 models ordinary generated text-token memory traffic; audio-code generation and waveform synthesis are outside this profile." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.162196680404945, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-thinker-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this mixed BF16/F32 repo. Activation traffic, multimodal encoder compute, Talker speech generation, Token2Wav waveform synthesis, and KV writes are outside Bounds Engine v1.", "notes": "The top-level and Thinker/Talker text configs record bfloat16 weights, while Token2Wav DiT/BigVGAN tensors are stored as F32 in the safetensors headers." }, "evidence": [ { "label": "Qwen2.5 Omni 3B API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Omni-3B", "source_type": "model_card", "supports": [ "repo", "pipeline", "license", "downloads", "safetensors_dtype_split", "commit_sha" ], "notes": "At commit f75b40e3da2003cdd6e1829b1f420ca70797c34e, the API reports an any-to-any Transformers repo with qwen2_5_omni tags, safetensors parameters BF16: 5088069376, F32: 449051296, total: 5537120672, and current downloads 1736821." }, { "label": "Qwen2.5 Omni 3B served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B/raw/f75b40e3da2003cdd6e1829b1f420ca70797c34e/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "thinker_talker_token2wav_scope" ], "notes": "The config records Qwen2_5OmniModel with enable_talker true and enable_audio_output true. The Thinker text config records hidden_size 2048, 36 layers, 16 attention heads, 2 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 32768, torch dtype BF16, and untied embeddings. The Talker config records 24 layers, hidden_size 896, 14 attention heads, and 2 KV heads. Token2Wav contains DiT and BigVGAN components, with DiT dtype F32." }, { "label": "Qwen2.5 Omni 3B model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B/blob/f75b40e3da2003cdd6e1829b1f420ca70797c34e/README.md", "source_type": "model_card", "supports": [ "ordinary_text_output_scope", "thinker_talker_token2wav_scope" ], "notes": "The model card describes Qwen2.5-Omni as a Thinker-Talker multimodal model that can generate text and natural speech. It documents disabling the Talker for text-only output as a memory-saving option and using return_audio false to return only text responses." }, { "label": "Qwen2.5 Omni 3B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B/raw/f75b40e3da2003cdd6e1829b1f420ca70797c34e/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "component_split" ], "notes": "The index records total_size 11972343936 bytes across three shards. Range-read shard headers found 2544 tensors totaling 11.972343936 GB: 10.176138752 GB BF16 tensors and 1.796205184 GB F32 tensors. Component totals are thinker.model 6.171877376 GB, thinker.lm_head 0.622329856 GB, thinker.audio_tower 1.275353088 GB, thinker.visual 1.337368576 GB, talker 0.769209856 GB, and token2wav 1.796205184 GB. thinker.model.embed_tokens.weight and thinker.lm_head.weight both have shape [151936, 2048] and contribute 0.622329856 GB each." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the model card, served config, safetensors index, and direct range-read shard header byte grouping. The ordinary text-decode scope follows the existing audited Qwen3-Omni Thinker/Talker profile convention." }, "notes": "Audited from HF API metadata, the served config, the model card, safetensors index metadata, and direct safetensors shard header range reads." }, { "id": "qwen--qwen2-5-omni-7b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Omni-7B-AWQ", "title": "Qwen2.5 Omni 7B AWQ 4-bit", "summary": "Audited memory-side bounds profile for ordinary text decode through the Qwen2.5-Omni 7B AWQ Thinker text decoder while charging the full Thinker/Talker/Token2Wav checkpoint as resident.", "model_family": "qwen2.5-omni-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-Omni-7B", "relation": "quantized", "source": "Hugging Face model card, served config comparison, and safetensors header grouping", "config_compatible": true, "notes": "The AWQ repo keeps the audited Qwen2.5-Omni 7B Thinker text geometry and context settings. The top-level architecture label changes to Qwen2_5OmniForConditionalGeneration and the artifact adds AWQ quantization metadata, while the Thinker text, Talker, and multimodal package scope match the BF16 base for the fields used by Bounds Engine v1." }, "architecture": { "canonical_architecture_id": "qwen2.5-omni-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.73222544, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 3.661606304, "resident_weight_gb": 12.70206784, "swept_weight_gb": 4.48075264, "auxiliary_resident_weight_gb": 8.2213152, "resident_parameter_scope": "logical Qwen2.5-Omni 7B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes thinker.model.embed_tokens.weight input lookup and includes Thinker text layer tensors, thinker.model.norm.weight, and thinker.lm_head.weight", "auxiliary_scope": "thinker.model.embed_tokens.weight, thinker.audio_tower tensors, thinker.visual tensors, all talker tensors, and all token2wav tensors are resident for the full Omni checkpoint but not swept as full matrices for each generated text token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, BF16 unquantized tensors, F16 scales/biases, and F32 Token2Wav tensors. The config records tie_word_embeddings false, and the headers store separate thinker.model.embed_tokens.weight and thinker.lm_head.weight tensors with identical shape. Ordinary text decode excludes the input embedding lookup and includes thinker.lm_head.weight." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Thinker text config records use_sliding_window false, so ordinary text decode charges full-context K and V streams for all 28 Thinker text layers. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "The full Qwen2.5-Omni AWQ checkpoint contains a Thinker text/vision/audio model, a Talker model, and Token2Wav waveform modules. Bounds Engine v1 models ordinary generated text-token memory traffic; audio-code generation and waveform synthesis are outside this profile." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-omni-thinker-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized BF16/F32 tensors from safetensors headers. AWQ dequantization, CPU offload, activation traffic, multimodal encoder compute, Talker speech generation, Token2Wav waveform synthesis, and KV writes are outside Bounds Engine v1.", "notes": "The config records BF16 Thinker and Talker dtypes plus AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert: [\"visual\"]. KV cache is charged at two bytes per scalar because the repo does not record a KV quantization scheme." }, "evidence": [ { "label": "Qwen2.5 Omni 7B AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Omni-7B-AWQ", "source_type": "model_card", "supports": [ "repo", "pipeline", "license", "downloads", "safetensors_dtype_split", "commit_sha", "weight_format" ], "notes": "At commit 848615c28dfd87b0f624baea4be1f29bec0d1db1, the API reports a public any-to-any Transformers repo with qwen2_5_omni, text-to-audio, multimodal, 4-bit, AWQ, and region:us tags; current downloads 125455; and safetensors logical parameters I32 6525288448, BF16 3757756672, F16 129024, F32 449051296, total 10732225440." }, { "label": "Qwen2.5 Omni 7B AWQ served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ/raw/848615c28dfd87b0f624baea4be1f29bec0d1db1/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "weight_format", "thinker_talker_token2wav_scope" ], "notes": "The config records Qwen2_5OmniForConditionalGeneration with enable_talker true and enable_audio_output true. The quantization config records AWQ 4-bit GEMM, group_size 128, zero_point true, and modules_to_not_convert visual. The Thinker text config records hidden_size 3584, 28 layers, 28 attention heads, 4 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 32768, BF16 dtype, and untied embeddings. The Talker config records 24 layers, hidden_size 896, 12 attention heads, 4 KV heads, and BF16 dtype." }, { "label": "Qwen2.5 Omni 7B AWQ model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ/blob/848615c28dfd87b0f624baea4be1f29bec0d1db1/README.md", "source_type": "model_card", "supports": [ "ordinary_text_output_scope", "thinker_talker_token2wav_scope", "weight_format" ], "notes": "The model card describes Qwen2.5-Omni as a Thinker-Talker multimodal model that can generate text and natural speech. It describes this repo as an AWQ low-VRAM variant with 4-bit Thinker weight quantization, on-demand loading/offload support, token2wav streaming inference, and return_audio false for text-only responses." }, { "label": "Qwen2.5 Omni 7B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ/raw/848615c28dfd87b0f624baea4be1f29bec0d1db1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "component_split", "weight_format" ], "notes": "The index records total_size 12702067840 bytes across four shards. Range-read shard headers found 2840 tensors totaling 12.702067840 GB: 7.515513344 GB BF16 tensors, 3.288133632 GB I32 packed tensors, 1.796205184 GB F32 tensors, and 0.102215680 GB F16 tensors. Component totals after separating the embedding and head are Thinker non-embedding text tensors 3.390757888 GB, thinker.model.embed_tokens.weight 1.089994752 GB, thinker.lm_head 1.089994752 GB, thinker.audio_tower 1.279294464 GB, thinker.visual 1.353100288 GB, talker 2.702720512 GB, and token2wav 1.796205184 GB. Ordinary swept text traffic is Thinker non-embedding text tensors plus lm_head, or 4.480752640 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, the model card, served config, safetensors index, and direct range-read shard header byte grouping. The ordinary text-decode scope follows the existing audited Qwen2.5-Omni 3B, Qwen2.5-Omni 7B, and Qwen3-Omni Thinker/Talker profile convention." }, "notes": "This profile models text-token decode with return_audio false while keeping the full AWQ checkpoint resident. A deployment with Talker disabled or CPU offload enabled should use a separate profile rather than silently reusing this one." }, { "id": "qwen--qwen2-5-omni-7b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-Omni-7B", "title": "Qwen2.5 Omni 7B", "summary": "Audited memory-side bounds profile for ordinary text decode through the Qwen2.5-Omni 7B Thinker text decoder while charging the full Thinker/Talker/Token2Wav checkpoint as resident.", "model_family": "qwen2.5-omni-dense", "architecture": { "canonical_architecture_id": "qwen2.5-omni-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.73222544, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 3.661606304, "resident_weight_gb": 22.362553472, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 8.2213152, "resident_parameter_scope": "safetensors_header_stored_bf16_f32_full_thinker_talker_token2wav_package", "swept_parameter_scope": "thinker.model text decoder excluding thinker.model.embed_tokens.weight plus thinker.lm_head.weight", "auxiliary_scope": "thinker.model.embed_tokens.weight, thinker.audio_tower tensors, thinker.visual tensors, all talker tensors, and all token2wav tensors are resident for the full Omni checkpoint but not swept as full matrices for each generated text token", "notes": "The config records tie_word_embeddings false, and the headers store separate thinker.model.embed_tokens.weight and thinker.lm_head.weight tensors with identical shape. Ordinary text decode excludes the input embedding lookup and includes thinker.lm_head.weight, so swept bytes equal the full thinker.model byte total. Talker and Token2Wav are retained in resident memory because the catalog row targets the full checkpoint; a talker-disabled deployment should be represented separately." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Thinker text config records use_sliding_window false, so ordinary text decode charges full-context K and V streams for all 28 Thinker text layers. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "The full Qwen2.5-Omni checkpoint contains a Thinker text/vision/audio model, a Talker model, and Token2Wav waveform modules. Bounds Engine v1 models ordinary generated text-token memory traffic; audio-code generation and waveform synthesis are outside this profile." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0836827922615835, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-thinker-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this mixed BF16/F32 repo. Activation traffic, multimodal encoder compute, Talker speech generation, Token2Wav waveform synthesis, and KV writes are outside Bounds Engine v1.", "notes": "The top-level, Thinker text, and Talker configs record bfloat16 weights, while Token2Wav tensors are stored as F32 in the safetensors headers." }, "evidence": [ { "label": "Qwen2.5 Omni 7B API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Omni-7B", "source_type": "model_card", "supports": [ "repo", "pipeline", "license", "downloads", "safetensors_dtype_split", "commit_sha" ], "notes": "At commit ae9e1690543ffd5c0221dc27f79834d0294cba00, the API reports a public any-to-any Transformers repo with qwen2_5_omni and multimodal tags, current downloads 614032, and safetensors parameters BF16 10283174144, F32 449051296, total 10732225440." }, { "label": "Qwen2.5 Omni 7B served config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B/raw/ae9e1690543ffd5c0221dc27f79834d0294cba00/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "thinker_talker_token2wav_scope" ], "notes": "The config records Qwen2_5OmniModel with enable_talker true and enable_audio_output true. The Thinker text config records hidden_size 3584, 28 layers, 28 attention heads, 4 KV heads, max_position_embeddings 32768, use_sliding_window false, sliding_window 32768, and untied embeddings. The Thinker audio config records a 32-layer audio encoder; the Thinker vision config records a 32-layer visual tower. The Talker config records 24 layers, hidden_size 896, 12 attention heads, 4 KV heads, and BF16 dtype. Token2Wav contains F32 waveform components." }, { "label": "Qwen2.5 Omni 7B model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B/blob/ae9e1690543ffd5c0221dc27f79834d0294cba00/README.md", "source_type": "model_card", "supports": [ "ordinary_text_output_scope", "thinker_talker_token2wav_scope" ], "notes": "The model card describes Qwen2.5-Omni as a Thinker-Talker multimodal model that can generate text and natural speech. It documents disabling the Talker for text-only output as a memory-saving option and using return_audio false to return only text responses." }, { "label": "Qwen2.5 Omni 7B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B/raw/ae9e1690543ffd5c0221dc27f79834d0294cba00/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "component_split" ], "notes": "The index records total_size 22366403936 bytes across five shards. Range-read shard headers found 2448 tensors totaling 22.362553472 GB: 20.566348288 GB BF16 tensors and 1.796205184 GB F32 tensors. Component totals are thinker.model 14.141238272 GB, thinker.lm_head 1.089994752 GB, thinker.model.embed_tokens.weight 1.089994752 GB, thinker.audio_tower 1.279294464 GB, thinker.visual 1.353100288 GB, talker 2.702720512 GB, and token2wav 1.796205184 GB. thinker.model.embed_tokens.weight and thinker.lm_head.weight both have shape [152064, 3584]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, served config, safetensors index, and direct range-read shard header byte grouping. The ordinary text-decode scope follows the existing audited Qwen2.5-Omni 3B and Qwen3-Omni Thinker/Talker profile convention." }, "notes": "This profile models text-token decode with return_audio false while keeping the full checkpoint resident. A deployment with Talker disabled should use a separate profile rather than silently reusing this one." }, { "id": "qwen--qwen2-5-vl-32b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-32B-Instruct", "title": "Qwen2.5 VL 32B Instruct BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the BF16/F32 Qwen2.5-VL 32B Instruct repo.", "model_family": "qwen2.5-vl-dense", "architecture": { "canonical_architecture_id": "qwen2-5-vl-32b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.452718336, "swept_params_b": 31.985308672, "auxiliary_resident_params_b": 1.467409664, "resident_weight_gb": 68.28312064, "swept_weight_gb": 63.970617344, "auxiliary_resident_weight_gb": 4.312503296, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model layers, model.norm.weight, and lm_head.weight", "auxiliary_scope": "visual tower tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each generated text token", "notes": "The swept subset includes model.* tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors and model.embed_tokens.weight. Text and lm_head tensors are BF16; visual tensors are F32." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0411830199914553, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16/F32 repo.", "notes": "The repo config records torch_dtype bfloat16. Direct safetensors header reads show BF16 text/lm_head/input-embedding tensors and F32 visual tower tensors; KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 VL 32B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-32B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "weight_format" ], "notes": "At repo SHA 7cfb30d71a1f4f49a57592323337a4a4727301da, the API records a public/non-gated Apache-2.0 image-text-to-text repo with transformers, safetensors, multimodal, eval-results, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 446745. The API safetensors block reports BF16: 32763876352, F32: 688841984, total: 33452718336. The model card states this repo contains the instruction-tuned 32B Qwen2.5-VL model and describes support for image, video, visual localization, and agentic visual workflows." }, { "label": "Qwen2.5 VL 32B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-32B-Instruct/raw/7cfb30d71a1f4f49a57592323337a4a4727301da/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, model_type qwen2_5_vl, torch_dtype bfloat16, tie_word_embeddings false, 64 text layers, hidden_size 5120, intermediate_size 27648, 40 attention heads, 8 KV heads, 128000 max position embeddings, use_sliding_window false, sliding_window 32768, mrope sections [16, 24, 24], and a resident visual config with hidden_size 1280, intermediate_size 3456, out_hidden_size 5120, and bfloat16 dtype. The card still contains an older sentence saying the current config is set for 32768 tokens, so this profile treats the pinned raw config as authoritative." }, { "label": "Qwen2.5 VL 32B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-32B-Instruct/resolve/7cfb30d71a1f4f49a57592323337a4a4727301da/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The index records total_size 68283120640 bytes across 18 shards. Range-read safetensors headers found 1161 tensors totaling 68.283120640 GB: BF16 65.527752704 GB and F32 2.755367936 GB. All 390 F32 tensors are under visual.*. model.embed_tokens.weight has shape [152064, 5120] and contributes 0.778567680B params / 1.557135360 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains swept. Ordinary text swept tensors, defined as model.* excluding embed_tokens plus lm_head, sum to 31.985308672B params / 63.970617344 GB. Auxiliary resident tensors, defined as visual.* plus model.embed_tokens.weight, sum to 1.467409664B params / 4.312503296 GB." }, { "label": "Qwen2.5 32B Instruct text profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "swept_weight_gb", "embedding_layout" ], "notes": "The swept text/logit subset is byte-identical to the already audited Qwen/Qwen2.5-32B-Instruct profile: 31.985308672B BF16 params / 63.970617344 GB, with an untied input embedding excluded from ordinary decode and a separate lm_head.weight included in swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, safetensors index, direct range-read safetensors shard headers, the existing Qwen2.5-VL profile convention, and byte comparison against the audited Qwen2.5 32B text profile." }, "notes": "This profile supersedes the generated metadata estimate, which treated all resident bytes as swept decode traffic and did not separate the F32 visual tower plus input embedding from ordinary text/logit decode traffic." }, { "id": "qwen--qwen2-5-vl-3b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-3B-Instruct-AWQ", "title": "Qwen2.5 VL 3B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the official AWQ 4-bit Qwen2.5-VL 3B Instruct repo.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-3B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served AWQ config, BF16 base config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-VL-3B-Instruct as its quantized base model. Manual comparison found matching text geometry and context fields; the AWQ artifact adds quantization_config and truncates vision_config metadata to hidden_size, while modules_to_not_convert visual plus direct BF16 visual tensors preserve the resident visual tower from the base profile." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-3b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.754622976, "swept_params_b": 3.085938688, "auxiliary_resident_params_b": 0.668684288, "resident_weight_gb": 3.401637888, "swept_weight_gb": 2.064269312, "auxiliary_resident_weight_gb": 1.337368576, "resident_parameter_scope": "logical Qwen2.5-VL 3B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode charges all model.* tensors, including model.embed_tokens.weight because the config ties embeddings and the safetensors header has no separate lm_head.weight", "auxiliary_scope": "visual tensors are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from the single safetensors header because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, BF16 unquantized tensors, BF16 embeddings, and BF16 visual tensors. Logical parameter counts follow the BF16 base/API model parameter count rather than treating AWQ scales and zeros as model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-multimodal-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, F16 scales/biases, and unquantized BF16 tensors from the safetensors header. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert: [\"visual\"]. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 VL 3B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-3B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit e7b623934290c5a4da0ee3c6e1e57bfb6b5abbf2, the API records a public non-gated image-text-to-text repo with transformers, safetensors, qwen2_5_vl, base_model Qwen/Qwen2.5-VL-3B-Instruct, text-generation-inference, endpoints_compatible, 4-bit, awq, region:us, qwen-research license metadata, and 176189 downloads. The API safetensors block reports logical parameters split across I32: 2774532096, BF16: 979998720, F16: 92160, total: 3754622976." }, { "label": "Qwen2.5 VL 3B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct-AWQ/raw/e7b623934290c5a4da0ee3c6e1e57bfb6b5abbf2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, bfloat16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert visual, 36 text layers, hidden size 2048, intermediate size 11008, 16 attention heads, 2 KV heads, 128000 max position embeddings, tie_word_embeddings true, use_sliding_window false, mRoPE, vocab size 151936, and a resident visual tower represented in config by hidden_size 1280." }, { "label": "Qwen2.5 VL 3B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "auxiliary_resident_scope" ], "notes": "Manual comparison found matching architecture, model_type, tie_word_embeddings, text layer count, hidden size, intermediate size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, rope_theta, and mRoPE section between the AWQ config and the BF16 Instruct config. The AWQ config adds quantization_config and omits detailed vision_config fields, but the base config records visual depth 32, hidden_size 1280, intermediate_size 3420, 16 heads, patch_size 14, spatial_merge_size 2, and full-attention blocks 7/15/23/31." }, { "label": "Qwen2.5 VL 3B Instruct AWQ safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct-AWQ/resolve/e7b623934290c5a4da0ee3c6e1e57bfb6b5abbf2/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The single safetensors file was range-read directly. The linked Content-Length is 3401785760 bytes, with a 147864-byte header and 3401637888 tensor bytes across 1328 tensors. Stored tensor bytes split into 504 I32 tensors / 1.398104064 GB, 464 BF16 tensors / 1.959997440 GB, and 360 F16 tensors / 0.043536384 GB. Text model tensors excluding embeddings are 1.441939456 GB; model.embed_tokens.weight is BF16 with shape [151936, 2048] and contributes 0.622329856 GB; visual.* tensors are 1.337368576 GB resident-only. The header has no lm_head.weight and the config ties embeddings, so model.embed_tokens.weight is charged as the swept output projection. Ordinary text swept tensors total 2.064269312 GB; resident-only visual tensors total 1.337368576 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, served AWQ config, BF16 Instruct base config comparison, existing BF16 3B and AWQ 7B profile conventions, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, visual tensors, and tied BF16 embedding/output traffic." }, { "id": "qwen--qwen2-5-vl-3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-3B-Instruct", "title": "Qwen2.5 VL 3B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2.5-VL 3B Instruct repo.", "model_family": "qwen2.5-vl-dense", "architecture": { "canonical_architecture_id": "qwen2-5-vl-3b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 3.754622976, "swept_params_b": 3.085938688, "auxiliary_resident_params_b": 0.668684288, "resident_weight_gb": 7.509245952, "swept_weight_gb": 6.171877376, "auxiliary_resident_weight_gb": 1.337368576, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model safetensors headers including tied embed_tokens output projection", "auxiliary_scope": "visual tower tensors are resident for the multimodal package but not swept for each generated text token", "notes": "The swept subset includes all model.* tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 VL 3B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct", "source_type": "model_card", "supports": [ "repo", "pipeline" ], "notes": "The scraped catalog and Hugging Face metadata identify an image-text-to-text multimodal repo." }, { "label": "Qwen2.5 VL 3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, tie_word_embeddings true, bfloat16, 36 text layers, 2 KV heads, hidden_size 2048, 16 attention heads, 128000 max position embeddings, use_sliding_window false, and a resident visual tower." }, { "label": "Qwen2.5 VL 3B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-3B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "pipeline" ], "notes": "At commit 66285546d2b821cf421d4f5eb2576359d3770cd3, the API safetensors block records BF16: 3754622976 and total: 3754622976, which this profile stores as 3.754622976B resident parameters." }, { "label": "Qwen2.5 VL 3B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists two safetensors shards. Range-read shard headers record 824 BF16 tensors totaling 3754622976 parameters and 7.509245952 GB, matching index total_size. model.* tensors sum to 3085938688 parameters / 6.171877376 GB, including model.embed_tokens.weight with shape [151936, 2048] and 311164928 parameters. The index has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. visual.* tensors sum to 668684288 parameters / 1.337368576 GB resident-only for ordinary text decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for text decode bounds. It deliberately separates resident visual weights from per-token swept language/logit weights." }, { "id": "qwen--qwen2-5-vl-72b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-72B-Instruct-AWQ", "title": "Qwen2.5 VL 72B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the official AWQ 4-bit Qwen2.5-VL 72B Instruct repo.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-72B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served AWQ config, BF16 base config comparison, and direct safetensors header metadata", "config_compatible": false, "notes": "The AWQ repo records Qwen/Qwen2.5-VL-72B-Instruct as its quantized base model. Manual comparison found matching architecture, context, layer count, hidden size, attention heads, KV heads, and vocabulary fields, but the served AWQ config pads intermediate_size to 29696 instead of the BF16 base's 29568 and truncates some vision_config metadata. This profile therefore uses the AWQ package's served config and exact stored tensor bytes." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-72b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 73.662435584, "swept_params_b": 71.7121536, "auxiliary_resident_params_b": 1.950281984, "resident_weight_gb": 43.004666368, "swept_weight_gb": 39.1041024, "auxiliary_resident_weight_gb": 3.900563968, "resident_parameter_scope": "logical Qwen2.5-VL 72B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and resident visual tensors while including layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "visual tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, BF16 unquantized tensors, F16 scales/biases, and BF16 visual, embedding, and head tensors. Logical parameter counts follow the Hugging Face API safetensors parameter total; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 8192 divided by 64 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-multimodal-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, F16 scales/biases, and unquantized BF16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert: [\"visual\"]. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 VL 72B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-72B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At commit c8b87d4b81f34b6a147577a310d7e75f0698f6c2, the API records a public non-gated Qwen-license image-text-to-text repo with transformers, safetensors, qwen2_5_vl, base_model Qwen/Qwen2.5-VL-72B-Instruct, text-generation-inference, endpoints_compatible, 4-bit, awq, and region:us metadata. The API safetensors block reports logical parameters split across I32 70464307200, BF16 3197309184, F16 819200, total 73662435584. Current API downloads were 184780 when audited." }, { "label": "Qwen2.5 VL 72B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct-AWQ/raw/c8b87d4b81f34b6a147577a310d7e75f0698f6c2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, bfloat16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert visual, 80 text layers, hidden size 8192, intermediate size 29696, 64 attention heads, 8 KV heads, 128000 max position embeddings, tie_word_embeddings false, use_sliding_window false, mRoPE, vocab size 152064, and a resident visual tower represented in config by hidden_size 1280, intermediate_size 3456, out_hidden_size 8192, and BF16 dtype." }, { "label": "Qwen2.5 VL 72B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/raw/89c86200743eec961a297729e7990e8f2ddbc4c5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "auxiliary_resident_scope" ], "notes": "Manual comparison found matching architecture, model_type, tie_word_embeddings, text layer count, hidden size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, and mRoPE section between the AWQ config and the BF16 Instruct config. Differences are intermediate_size 29696 in AWQ versus 29568 in BF16, quantization_config, rope_scaling label spelling, and truncated AWQ vision_config metadata." }, { "label": "Qwen2.5 VL 72B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct-AWQ/raw/c8b87d4b81f34b6a147577a310d7e75f0698f6c2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 43004666368 bytes across 11 shards. Direct range-read safetensors headers found 2473 tensors totaling the same 43.004666368 GB: 1120 I32 tensors / 35.507404800 GB, 553 BF16 tensors / 6.394618368 GB, and 800 F16 tensors / 1.102643200 GB. model.embed_tokens.weight is BF16 with shape [152064, 8192] and contributes 1.245708288 logical parameters / 2.491416576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. visual.* tensors total 0.704573696 logical parameters / 1.409147392 GB resident-only. Layer tensors plus model.norm.weight plus lm_head.weight total 39.104102400 GB swept traffic. Auxiliary resident tensors total 3.900563968 GB." }, { "label": "Qwen2.5 VL 72B Instruct BF16 profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/raw/89c86200743eec961a297729e7990e8f2ddbc4c5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "base_model_proof", "embedding_layout", "auxiliary_resident_scope" ], "notes": "The audited BF16 base profile uses the same untied embedding layout: model.embed_tokens.weight is resident-only for ordinary decode and separate lm_head.weight is swept. The AWQ profile preserves that layout while changing stored bytes and padding intermediate_size in the served config." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served AWQ config, BF16 Instruct base config comparison, existing BF16 72B and AWQ 3B/7B profile conventions, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ qzeros/scales, BF16 visual tensors, and BF16 embedding/output traffic. It intentionally uses exact package bytes rather than inheriting the BF16 72B swept traffic because the AWQ config pads the MLP intermediate size." }, { "id": "qwen--qwen2-5-vl-72b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-72B-Instruct", "title": "Qwen2.5 VL 72B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2.5-VL 72B Instruct repo.", "model_family": "qwen2.5-vl-dense", "architecture": { "canonical_architecture_id": "qwen2-5-vl-72b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 73.410777344, "swept_params_b": 71.46049536, "auxiliary_resident_params_b": 1.950281984, "resident_weight_gb": 146.821554688, "swept_weight_gb": 142.92099072, "auxiliary_resident_weight_gb": 3.900563968, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model layers, model.norm.weight, and lm_head.weight", "auxiliary_scope": "visual tower tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each generated text token", "notes": "The swept subset includes model.* tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors and model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 80, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 8192 divided by 64 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and direct safetensors header reads show only BF16 tensors. KV cache is charged at BF16 two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 VL 72B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-72B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "total_params_b", "weight_format" ], "notes": "At repo SHA 89c86200743eec961a297729e7990e8f2ddbc4c5, the API records a public/non-gated Qwen-license image-text-to-text repo with transformers, safetensors, qwen2_5_vl, multimodal, eval-results, text-generation-inference, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 389971. The API safetensors block reports BF16: 73410777344 and total: 73410777344. The model card states this repo contains the instruction-tuned 72B Qwen2.5-VL model and describes support for images, video, visual localization, and agentic visual workflows." }, { "label": "Qwen2.5 VL 72B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/raw/89c86200743eec961a297729e7990e8f2ddbc4c5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, model_type qwen2_5_vl, torch_dtype bfloat16, tie_word_embeddings false, 80 text layers, hidden_size 8192, intermediate_size 29568, 64 attention heads, 8 KV heads, 128000 max position embeddings, use_sliding_window false, sliding_window 32768, mrope sections [16, 24, 24], and a resident visual config with hidden_size 1280, intermediate_size 3456, out_hidden_size 8192, 32 vision blocks, and BF16 dtype. The card still contains an older sentence saying the current config is set for 32768 tokens, so this profile treats the pinned raw config as authoritative." }, { "label": "Qwen2.5 VL 72B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/resolve/89c86200743eec961a297729e7990e8f2ddbc4c5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The index records total_size 146821554688 bytes across 38 shards. Range-read safetensors headers found 1353 BF16 tensors totaling 73.410777344B params / 146.821554688 GB, matching index total_size. model.embed_tokens.weight has shape [152064, 8192] and contributes 1.245708288B params / 2.491416576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains swept. Ordinary text swept tensors, defined as model.* excluding embed_tokens plus lm_head, sum to 71.460495360B params / 142.920990720 GB. Auxiliary resident tensors, defined as visual.* plus model.embed_tokens.weight, sum to 1.950281984B params / 3.900563968 GB." }, { "label": "Qwen2.5 72B Instruct text profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "swept_weight_gb", "embedding_layout" ], "notes": "The swept text/logit subset is byte-identical to the already audited Qwen/Qwen2.5-72B-Instruct profile: 71.460495360B BF16 params / 142.920990720 GB, with an untied input embedding excluded from ordinary decode and a separate lm_head.weight included in swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, safetensors index, direct range-read safetensors shard headers, the existing Qwen2.5-VL profile convention, and comparison against the audited Qwen2.5 72B text profile." }, "notes": "This profile supersedes the generated metadata estimate, which treated all resident bytes as swept decode traffic and did not separate the visual tower plus input embedding from ordinary text/logit decode traffic." }, { "id": "qwen--qwen2-5-vl-7b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-7B-Instruct-AWQ", "title": "Qwen2.5 VL 7B Instruct AWQ 4-bit", "summary": "Audited memory-side text-decode bounds profile for the official AWQ 4-bit Qwen2.5-VL 7B Instruct repo.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-7B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2.5-VL-7B-Instruct as its quantized base model. Manual config comparison found matching text geometry and context fields; the AWQ artifact adds quantization_config and keeps the text decode geometry of the BF16 base repo." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-7b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.292166656, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 1.22154752, "resident_weight_gb": 6.92384768, "swept_weight_gb": 4.48075264, "auxiliary_resident_weight_gb": 2.44309504, "resident_parameter_scope": "logical Qwen2.5-VL 7B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and resident visual tensors while including layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "visual tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, BF16 unquantized tensors, F16 scales/biases, and BF16 visual, embedding, and head tensors. Logical parameter counts match the BF16 base model; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-multimodal-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized BF16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert: [\"visual\"]. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2.5 VL 7B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-7B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 536a35794df8831aa814970ee8f89eff577e7718, the API records an Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen2.5-VL-7B-Instruct and tags 4-bit and awq. The API safetensors block reports logical parameters split across I32: 6525288448, BF16: 1766749184, F16: 129024, total: 8292166656. The card says this repo contains the instruction-tuned 7B Qwen2.5-VL model with AWQ and says the current config is set for context length up to 32768 tokens, while the raw config records 128000 max_position_embeddings." }, { "label": "Qwen2.5 VL 7B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-AWQ/raw/536a35794df8831aa814970ee8f89eff577e7718/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, bfloat16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert visual, 28 text layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, 128000 max position embeddings, tie_word_embeddings false, use_sliding_window false, mrope, vocab size 152064, and a resident visual tower." }, { "label": "Qwen2.5 VL 7B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/cc594898137f460bfe9f0759e9844b3ce807cfb5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching architecture, model_type, tie_word_embeddings, text layer count, hidden size, intermediate size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, and mRoPE section between the AWQ config and the BF16 Instruct config. Differences are quantization_config, _name_or_path, transformers_version, rope_scaling label spelling, and truncated AWQ vision_config metadata." }, { "label": "Qwen2.5 VL 7B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-AWQ/raw/536a35794df8831aa814970ee8f89eff577e7718/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 6923847680 bytes across two shards. Range-read safetensors headers found 1121 tensors totaling 6.92384768 GB: 3.288133632 GB I32 packed tensors, 3.533498368 GB BF16 tensors, and 0.10221568 GB F16 tensors. model.embed_tokens.weight is BF16 with shape [152064, 3584] and contributes 544997376 logical parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. visual.* tensors total 1.353100288 GB resident-only. Layer tensors plus model.norm.weight plus lm_head.weight total 4.48075264 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served AWQ config, BF16 Instruct config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, and unquantized visual, embedding, and head tensors." }, { "id": "qwen--qwen2-5-vl-7b-instruct", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2.5-VL-7B-Instruct", "title": "Qwen2.5 VL 7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2.5-VL 7B Instruct repo.", "model_family": "qwen2.5-vl-dense", "architecture": { "canonical_architecture_id": "qwen2-5-vl-7b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.292166656, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 1.22154752, "resident_weight_gb": 16.584333312, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 2.44309504, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "visual tower tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each generated text token", "notes": "The swept subset includes model.* tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors and model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and safetensors headers record BF16 tensors." }, "evidence": [ { "label": "Qwen2.5 VL 7B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline" ], "notes": "The model metadata identifies an Apache-2.0 image-text-to-text multimodal repo." }, { "label": "Qwen2.5 VL 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, tie_word_embeddings false, bfloat16, 28 text layers, 4 KV heads, hidden_size 3584, 28 attention heads, 128000 max position embeddings, use_sliding_window false, and a resident visual tower." }, { "label": "Qwen2.5 VL 7B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-7B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit cc594898137f460bfe9f0759e9844b3ce807cfb5 records safetensors parameters BF16: 8292166656 and total: 8292166656." }, { "label": "Qwen2.5 VL 7B Instruct safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all five shards. Stored tensors sum to 8292166656 BF16 params and 16.584333312 GB. model.embed_tokens.weight is 544997376 BF16 params and 1.089994752 GB. lm_head.weight is a separate untied tensor of the same size. Ordinary text swept tensors, defined as model.* excluding embed_tokens plus lm_head, sum to 7070619136 BF16 params and 14.141238272 GB. Auxiliary resident tensors, defined as visual.* plus model.embed_tokens.weight, sum to 1221547520 BF16 params and 2.44309504 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, HF API metadata, and safetensors header parameter split." }, "notes": "This profile is for text decode bounds. It deliberately separates resident visual/input-embedding weights from per-token swept language/logit weights." }, { "id": "qwen--qwen2-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-7B-Instruct", "title": "Qwen2 7B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen2 7B Instruct repo.", "model_family": "qwen2-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-7B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the Instruct repo records 32768 max position embeddings while the base config records 131072. This profile uses the Instruct repo config directly." }, "architecture": { "canonical_architecture_id": "qwen2-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "notes": "Range-read safetensors headers record 7615616512 BF16 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Dense Qwen2ForCausalLM profile using the served Instruct repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2 7B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "total_params_b", "weight_format" ], "notes": "At commit f2826a00ceef68f0f2b946d945ecc0477ce4450c, the API reports a public Apache-2.0 text-generation repo with base_model Qwen/Qwen2-7B, region:us, current downloads 692373, and BF16 safetensors count 7615616512." }, { "label": "Qwen2 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/raw/f2826a00ceef68f0f2b946d945ecc0477ce4450c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen2ForCausalLM, bfloat16, 28 layers, hidden_size 3584, intermediate_size 18944, 28 attention heads, 4 KV heads, 32768 max position embeddings, sliding_window 131072, use_sliding_window false, tie_word_embeddings false, vocab_size 152064, and rope_theta 1000000." }, { "label": "Qwen2 7B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/resolve/f2826a00ceef68f0f2b946d945ecc0477ce4450c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "weight_format", "embedding_layout" ], "notes": "The index lists four safetensors shards with total_size 15231233024 bytes. Range-read shard headers record 339 BF16 tensors totaling 7615616512 parameters and 15.231233024 GB, matching the index. model.embed_tokens.weight has shape [152064, 3584] and contributes 544997376 parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7070619136 parameters / 14.141238272 GB." }, { "label": "Qwen2 7B base config", "url": "https://huggingface.co/Qwen/Qwen2-7B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but a different max_position_embeddings value, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the served config, base config comparison, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen2-audio-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-Audio-7B-Instruct", "title": "Qwen2 Audio 7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2 Audio 7B Instruct repo.", "model_family": "qwen2-audio-dense", "architecture": { "canonical_architecture_id": "qwen2-audio-7b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.397094912, "swept_params_b": 7.115771904, "auxiliary_resident_params_b": 1.281323008, "resident_weight_gb": 16.794189824, "swept_weight_gb": 14.231543808, "auxiliary_resident_weight_gb": 2.562646016, "resident_parameter_scope": "safetensors_header_stored_bf16_full_qwen2_audio_package", "swept_parameter_scope": "ordinary text decode excludes audio_tower, multi_modal_projector, and language_model.model.embed_tokens.weight input lookup; it includes language_model.lm_head.weight output projection", "auxiliary_scope": "audio_tower and multi_modal_projector are resident for audio prefill, and the input embedding table is resident lookup traffic, but none are swept as full matrices per generated text token", "notes": "Range-read safetensors headers record 8397094912 BF16 stored parameters. The checkpoint stores the audio encoder, multimodal projector, Qwen2 language model, separate input embedding, and separate lm_head tensors. This profile charges the full package as resident and charges only language_model tensors other than the input embedding as ordinary text-decode swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 32, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text_config is a sparse Qwen2Config override. Transformers Qwen2Config defaults provide 32 layers, 32 attention heads, 32 KV heads, hidden size 4096, and use_sliding_window false. Header tensor shapes confirm 32 language layers with q/k/v/o projection shapes [4096, 4096], so the ordinary text-decode cache is full-context K/V with 32 KV heads and 128 head dimension." }, "notes": "Qwen2AudioForConditionalGeneration combines a Whisper-style audio encoder, multimodal projector, and Qwen2 language model. This profile models ordinary text-token decode after any audio feature extraction and prefill, not audio encoder or audio prefill throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-audio-prefill-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo API records BF16 safetensors, and direct shard header reads found only BF16 tensors." }, "evidence": [ { "label": "Qwen2 Audio 7B Instruct HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-Audio-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit 0a095220c30b7b31434169c3086508ef3ea5bf0a, the API reports a public Apache-2.0 audio-text-to-text repo with qwen2_audio tags, region:us, current downloads 581582, and BF16 safetensors count 8397094912." }, { "label": "Qwen2 Audio 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct/raw/0a095220c30b7b31434169c3086508ef3ea5bf0a/config.json", "source_type": "config", "supports": [ "model_family", "audio_encoder", "max_context_tokens", "text_config_overrides", "audio_token_index", "serving" ], "notes": "The config records Qwen2AudioForConditionalGeneration, a qwen2_audio_encoder audio tower with 32 encoder layers, 20 audio attention heads, d_model 1280, 128 mel bins, and max_source_positions 1500. The text_config records Qwen2, bfloat16, 8192 max_position_embeddings, 32768 sliding_window, 11008 intermediate size, 156032 vocabulary, rope_theta 10000, and audio_token_index 151646." }, { "label": "Transformers Qwen2Config defaults", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.38.1/src/transformers/models/qwen2/configuration_qwen2.py", "source_type": "manual_review", "supports": [ "layers", "kv_heads", "head_dim", "kv_adapter", "sliding_window_boundary" ], "notes": "The Qwen2Config used by the served sparse text_config defaults to hidden_size 4096, 32 layers, 32 attention heads, 32 KV heads, tie_word_embeddings false, and use_sliding_window false. The model's text_config overrides context, intermediate size, rope, dtype, and vocabulary but does not override these core Qwen2 fields." }, { "label": "Transformers Qwen2 Audio implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/main/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py", "source_type": "manual_review", "supports": [ "ordinary_text_decode_scope", "audio_prefill_scope", "language_model_decode_path" ], "notes": "Manual review found Qwen2AudioModel embeds input tokens, computes audio features through the audio tower and projector, scatters those features into inputs_embeds at audio-token positions, and then calls self.language_model. Qwen2AudioForConditionalGeneration applies lm_head to the language-model hidden states. prepare_inputs_for_generation passes input_features only on the first cached iteration, so cached decode uses the language model path without re-running the audio tower." }, { "label": "Qwen2 Audio 7B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct/resolve/0a095220c30b7b31434169c3086508ef3ea5bf0a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout", "audio_encoder", "kv_adapter" ], "notes": "The index lists five safetensors shards with total_size 16794189824 bytes. Range-read shard headers record 876 BF16 tensors totaling 8397094912 parameters / 16.794189824 GB. audio_tower tensors sum to 636968960 params / 1.27393792 GB, multi_modal_projector tensors sum to 5246976 params / 0.010493952 GB, and language_model tensors sum to 7754878976 params / 15.509757952 GB. language_model.model.embed_tokens.weight has shape [156032, 4096] and contributes 639107072 params / 1.278214144 GB resident-only for ordinary decode. language_model.lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Language layers plus model.norm plus lm_head total 7115771904 params / 14.231543808 GB. Headers confirm 32 language layers; layer-0 q_proj/k_proj/v_proj/o_proj shapes are [4096, 4096], and gate/up/down MLP shapes are [11008, 4096], [11008, 4096], and [4096, 11008]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, Transformers Qwen2/Qwen2-Audio source, direct safetensors header grouping, and local scrape row." }, "notes": "This profile is for ordinary text-output decode after audio prefill. It does not estimate audio encoder throughput, audio feature extraction cost, or multimodal prefill latency." }, { "id": "qwen--qwen2-vl-2b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-VL-2B-Instruct", "title": "Qwen2 VL 2B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2-VL 2B Instruct repo.", "model_family": "qwen2-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-VL-2B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "The Instruct model card records Qwen/Qwen2-VL-2B as the base model. Manual config comparison found matching text and vision geometry for the fields used by this profile." }, "architecture": { "canonical_architecture_id": "qwen2-vl-2b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.2089856, "swept_params_b": 1.543714304, "auxiliary_resident_params_b": 0.665271296, "resident_weight_gb": 4.4179712, "swept_weight_gb": 3.087428608, "auxiliary_resident_weight_gb": 1.330542592, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model safetensors headers including tied embed_tokens output projection", "auxiliary_scope": "visual tensors are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "The swept subset includes all model.* tensors. The config records tie_word_embeddings true and the safetensors headers have no separate lm_head.weight, so model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 1536 divided by 12 attention heads." }, "notes": "Qwen2VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2 VL 2B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The card identifies this as the instruction-tuned 2B Qwen2-VL model, records Apache-2.0 licensing, identifies image-text-to-text multimodal usage, and lists Qwen/Qwen2-VL-2B as the base model." }, { "label": "Qwen2 VL 2B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/raw/895c3a49bc3fa70a340399125c650a463535e71c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2VLForConditionalGeneration, tie_word_embeddings true, bfloat16, 28 text layers, hidden size 1536, 12 attention heads, 2 KV heads, 32768 max position embeddings, use_sliding_window false, mrope, and a resident 32-layer visual tower." }, { "label": "Qwen2 VL 2B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-VL-2B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit 895c3a49bc3fa70a340399125c650a463535e71c, the API safetensors block records BF16: 2208985600 and total: 2208985600, which this profile stores as 2.2089856B resident parameters." }, { "label": "Qwen2 VL 2B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/raw/895c3a49bc3fa70a340399125c650a463535e71c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists two safetensors shards with metadata total_size 4417971200 bytes. Range-read shard headers record 729 BF16 tensors totaling 2208985600 parameters and 4.4179712 GB, matching index total_size. model.* tensors sum to 1543714304 parameters / 3.087428608 GB, including model.embed_tokens.weight with shape [151936, 1536] and 233373696 parameters / 0.466747392 GB. The headers have no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. visual.* tensors sum to 665271296 parameters / 1.330542592 GB resident-only for ordinary text decode." }, { "label": "Qwen2 VL 2B base config", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching text geometry, max context, tie_word_embeddings, mrope settings, and vision tower geometry for the fields used by this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, model card, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual weights from per-token swept language/logit weights while keeping the tied text embedding in swept traffic." }, { "id": "qwen--qwen2-vl-7b-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-VL-7B-Instruct-AWQ", "title": "Qwen2 VL 7B Instruct AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen2-VL 7B Instruct repo.", "model_family": "qwen2-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-VL-7B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata and served config comparison", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen2-VL-7B-Instruct as its quantized base model. Manual config comparison found matching Qwen2VLForConditionalGeneration text and vision geometry; the AWQ artifact adds quantization_config and changes torch_dtype from bfloat16 to float16." }, "architecture": { "canonical_architecture_id": "qwen2-vl-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.291375616, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 1.22075648, "resident_weight_gb": 6.9222656, "swept_weight_gb": 4.48075264, "auxiliary_resident_weight_gb": 2.44151296, "resident_parameter_scope": "logical Qwen2-VL 7B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and resident visual tensors while including layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "visual tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 tensors, F16 scales/zeros/biases, and unquantized F16 visual, embedding, and head tensors. Logical parameter counts match the BF16 base model; exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Qwen2VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-multimodal-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales/biases, and unquantized F16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert: [\"visual\"]. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen2 VL 7B Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-VL-7B-Instruct-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "At commit 6ec2560b0afc3a618d4acc9b8e2967d1642f463d, the API records an Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen2-VL-7B-Instruct and tags 4-bit and awq. The API safetensors block reports logical parameters split across I32: 6525288448 and F16: 1766087168, total 8291375616." }, { "label": "Qwen2 VL 7B Instruct AWQ config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct-AWQ/raw/6ec2560b0afc3a618d4acc9b8e2967d1642f463d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope" ], "notes": "The config records Qwen2VLForConditionalGeneration, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert visual, 28 text layers, hidden size 3584, 28 attention heads, 4 KV heads, 32768 max position embeddings, tie_word_embeddings false, use_sliding_window false, mrope, and a resident 32-layer visual tower." }, { "label": "Qwen2 VL 7B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/raw/eed13092ef92e448dd6875b2a00151bd3f7db0ac/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no audited geometry differences between the AWQ config and the BF16 Instruct config after excluding quantization_config, torch_dtype, _name_or_path, transformers_version, and JSON object key order inside rope_scaling." }, { "label": "Qwen2 VL 7B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct-AWQ/raw/6ec2560b0afc3a618d4acc9b8e2967d1642f463d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 6922265600 bytes across two shards. Range-read safetensors headers found 1122 tensors totaling 6.9222656 GB: 3.288133632 GB I32 packed tensors and 3.634131968 GB F16 tensors. model.embed_tokens.weight is F16 with shape [152064, 3584] and contributes 544997376 logical parameters / 1.089994752 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. visual.* tensors total 1.351518208 GB resident-only. Layer tensors plus model.norm.weight plus lm_head.weight total 4.48075264 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served AWQ config, BF16 Instruct config comparison, safetensors index, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, biases, and unquantized visual, embedding, and head tensors." }, { "id": "qwen--qwen2-vl-7b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen2-VL-7B-Instruct", "title": "Qwen2 VL 7B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen2-VL 7B Instruct repo.", "model_family": "qwen2-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2-VL-7B", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "The Instruct model card records Qwen/Qwen2-VL-7B as the base model. Manual config comparison found matching text and vision geometry for the fields used by this profile." }, "architecture": { "canonical_architecture_id": "qwen2-vl-7b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.291375616, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 1.22075648, "resident_weight_gb": 16.582751232, "swept_weight_gb": 14.141238272, "auxiliary_resident_weight_gb": 2.44151296, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "visual tower tensors and model.embed_tokens.weight are resident for the multimodal package but not swept for each generated text token", "notes": "The swept subset includes model.* tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual.* tensors and model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 3584 divided by 28 attention heads." }, "notes": "Qwen2VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen2 VL 7B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The card identifies this as the instruction-tuned 7B Qwen2-VL model, records Apache-2.0 licensing, identifies image-text-to-text multimodal usage, and lists Qwen/Qwen2-VL-7B as the base model." }, { "label": "Qwen2 VL 7B Instruct config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/raw/eed13092ef92e448dd6875b2a00151bd3f7db0ac/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2VLForConditionalGeneration, tie_word_embeddings false, bfloat16, 28 text layers, hidden size 3584, 28 attention heads, 4 KV heads, 32768 max position embeddings, use_sliding_window false, mrope, and a resident 32-layer visual tower." }, { "label": "Qwen2 VL 7B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-VL-7B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit eed13092ef92e448dd6875b2a00151bd3f7db0ac, the API safetensors block records BF16: 8291375616 and total: 8291375616, which this profile stores as 8.291375616B resident parameters." }, { "label": "Qwen2 VL 7B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/raw/eed13092ef92e448dd6875b2a00151bd3f7db0ac/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists five safetensors shards with metadata total_size 16582751232 bytes. Range-read shard headers record 730 BF16 tensors totaling 8291375616 parameters and 16.582751232 GB, matching index total_size. model.* tensors excluding model.embed_tokens.weight plus the separate lm_head.weight sum to 7070619136 parameters / 14.141238272 GB. model.embed_tokens.weight and lm_head.weight each have shape [152064, 3584] and 544997376 parameters / 1.089994752 GB. The config records tie_word_embeddings false, so lm_head.weight is the untied output projection and stays in swept decode traffic. visual.* tensors plus model.embed_tokens.weight sum to 1220756480 parameters / 2.44151296 GB resident-only for ordinary text decode." }, { "label": "Qwen2 VL 7B base config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching text geometry, max context, tie_word_embeddings, mrope settings, and vision tower geometry for the fields used by this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, model card, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/input-embedding weights from per-token swept language/logit weights while keeping the untied output projection in swept traffic." }, { "id": "qwen--qwen3-0-6b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-0.6B-Base", "title": "Qwen3 0.6B Base BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3 0.6B Base repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B-Base", "relation": "base", "source": "Hugging Face model metadata, served config, and direct safetensors header metadata", "config_compatible": true, "notes": "This repo is the canonical Qwen3 0.6B base checkpoint. Manual comparison with the already audited Qwen/Qwen3-0.6B fine-tune found matching core tensor geometry but a different context length: this base config records 32768 max position embeddings while the fine-tune records 40960." }, "architecture": { "canonical_architecture_id": "qwen3-0-6b-base", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.59604992, "swept_params_b": 0.59604992, "auxiliary_resident_params_b": 0, "resident_weight_gb": 1.19209984, "swept_weight_gb": 1.19209984, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode charges all stored BF16 tensors; model.embed_tokens.weight is charged as the tied output projection because no separate lm_head.weight tensor is stored", "auxiliary_scope": "No resident-only weight tensor is excluded from ordinary text decode in Bounds Engine v1.", "notes": "Range-read safetensors headers record 310 BF16 tensors totaling 596049920 stored parameters. model.embed_tokens.weight has shape [151936, 1024] and contributes 0.311164928 GB. The config marks tie_word_embeddings true and the checkpoint has no lm_head.weight, so the embedding table is also the output projection and remains in swept decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM base profile using the served repo config and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, the HF API safetensors metadata records only BF16 parameters, and the direct safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen3 0.6B Base model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model card/API metadata records a public Apache-2.0 text-generation repo with transformers, safetensors, qwen3, text-generation-inference, endpoints_compatible, region:us, and conversational tags." }, { "label": "Qwen3 0.6B Base HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-0.6B-Base", "source_type": "derived_calculation", "supports": [ "downloads", "total_params_b", "weight_format" ], "notes": "At commit da87bfb608c14b7cf20ba1ce41287e8de496c0cd, the API records 950256 downloads, safetensors parameters BF16 596049920, total 596049920, and usedStorage 1192135096." }, { "label": "Qwen3 0.6B Base config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, tie_word_embeddings true, hidden size 1024, intermediate size 3072, 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, rope_theta 1000000, vocab size 151936, and no sliding_window setting." }, { "label": "Qwen3 0.6B Base safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 35248-byte header with 310 BF16 tensors totaling 596049920 parameters / 1.19209984 GB. model.embed_tokens.weight has shape [151936, 1024] and contributes 0.311164928 GB. The file has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. Layer tensors plus model.norm.weight total 0.880934912 GB. The linked safetensors object size is 1192135096 bytes, exactly payload plus the 8-byte safetensors prefix and header JSON." }, { "label": "Qwen3 0.6B fine-tune comparison", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited Qwen/Qwen3-0.6B profile found matching layers, attention heads, KV heads, head dimension, hidden size, intermediate size, rope_theta, and vocab size. The base config records 32768 max position embeddings while the fine-tune records 40960." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, direct range-read safetensors header, existing Qwen3 0.6B profile comparison, and local scrape row." }, "notes": "This is a self-contained dense BF16 base-model profile for production profile-backed bounds. It does not inherit the Qwen/Qwen3-0.6B fine-tune context setting." }, { "id": "qwen--qwen3-0-6b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-0.6B-FP8", "title": "Qwen3 0.6B FP8", "summary": "Audited memory-side text-decode bounds profile for the FP8 Qwen3 0.6B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B", "relation": "quantized", "source": "Hugging Face model metadata, served config, base profile comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-0.6B as its base model and preserves the served Qwen3ForCausalLM tensor geometry: 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, 151936 vocab size, and 40960 max position embeddings. It adds fine-grained FP8 quantization with e4m3 format and 128x128 weight blocks." }, "architecture": { "canonical_architecture_id": "qwen3-0-6b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.751659264, "swept_params_b": 0.5960768, "auxiliary_resident_params_b": 0.155582464, "resident_weight_gb": 1.062916608, "swept_weight_gb": 0.75175168, "auxiliary_resident_weight_gb": 0.311164928, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 507 tensors totaling 751659264 stored parameters / 1.062916608 GB. The config marks tie_word_embeddings true, but the checkpoint stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic. Layer tensors are mostly F8_E4M3 with BF16 scale-inverse tensors." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting or FP8 KV cache scheme, so the v1 profile charges full-context BF16 K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.4140936710386902, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and BF16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 with quantization_config fp8/e4m3 dynamic activation scheme and 128x128 weight blocks. KV cache bytes are charged as BF16 because the config does not define a quantized KV cache." }, "evidence": [ { "label": "Qwen3 0.6B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-0.6B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit e5be08033360965ceca7b0ffd72d521a51331ce0, the API reports an Apache-2.0 text-generation repo with base_model Qwen/Qwen3-0.6B and safetensors parameters BF16: 311257344, F8_E4M3: 440401920, total: 751659264. The model card identifies this as the FP8 version of Qwen3-0.6B and documents fine-grained FP8 quantization with 128 block size." }, { "label": "Qwen3 0.6B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-FP8/raw/e5be08033360965ceca7b0ffd72d521a51331ce0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, 28 layers, hidden_size 1024, 16 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, use_sliding_window false, tie_word_embeddings true, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 0.6B base profile", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout" ], "notes": "Manual comparison with the audited BF16 Qwen3-0.6B profile found matching served architecture geometry and the same stored separate model.embed_tokens.weight plus lm_head.weight layout." }, { "label": "Qwen3 0.6B FP8 safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-FP8/resolve/e5be08033360965ceca7b0ffd72d521a51331ce0/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 59128-byte header with 507 tensors totaling 751659264 parameters / 1.062916608 GB: 440401920 F8_E4M3 parameters / 0.44040192 GB plus 311257344 BF16 parameters / 0.622514688 GB. model.embed_tokens.weight has shape [151936, 1024] and contributes 0.311164928 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 0.75175168 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card, served config, base profile comparison, direct safetensors header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-0-6b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-0.6B", "title": "Qwen3 0.6B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 0.6B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the served repo records 40960 max position embeddings while the base config records 32768. This profile therefore uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "qwen3-0-6b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.751632384, "swept_params_b": 0.59604992, "auxiliary_resident_params_b": 0.155582464, "resident_weight_gb": 1.503264768, "swept_weight_gb": 1.19209984, "auxiliary_resident_weight_gb": 0.311164928, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 311 BF16 tensors totaling 751632384 stored parameters. The config marks tie_word_embeddings true, but the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and direct safetensors headers record BF16 tensors." }, "evidence": [ { "label": "Qwen3 0.6B model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen3-0.6B-Base as the base model and Apache-2.0 as the license." }, { "label": "Qwen3 0.6B config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 28 layers, 8 KV heads, 128 head dimension, and 40960 max position embeddings." }, { "label": "Qwen3 0.6B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-0.6B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format" ], "notes": "At commit c1899de289a04d12100db370d81485cdf75e47ca, the API safetensors block records BF16: 751632384 and total: 751632384." }, { "label": "Qwen3 0.6B safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-0.6B/resolve/c1899de289a04d12100db370d81485cdf75e47ca/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 35552-byte header with 311 BF16 tensors totaling 751632384 parameters / 1.503264768 GB. model.embed_tokens.weight has shape [151936, 1024] and contributes 0.311164928 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 1.19209984 GB." }, { "label": "Qwen3 0.6B Base config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but a different max_position_embeddings value, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-1-7b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-1.7B-Base", "title": "Qwen3 1.7B Base BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 1.7B Base repo.", "model_family": "qwen3-dense", "architecture": { "canonical_architecture_id": "qwen3-1-7b-base", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 1.720574976, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.311164928, "non_embedding_params_b": 1.409410048, "notes": "A range-read of model.safetensors records 1720574976 BF16 stored parameters. model.embed_tokens.weight contributes 311164928 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records no sliding_window or attention_chunk_size setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM base-pretraining profile using the served base repo config." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and the model.safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen3 1.7B Base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-1.7B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit ea980cb0a6c2ae4b936e82123acc929f1cec04c1, the API records an Apache-2.0 text-generation repo with transformers, safetensors, qwen3, text-generation-inference, endpoints_compatible, region:us, 670780 downloads, and safetensors parameters BF16: 1720574976, total: 1720574976." }, { "label": "Qwen3 1.7B Base config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base/raw/ea980cb0a6c2ae4b936e82123acc929f1cec04c1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM, qwen3, bfloat16, hidden_size 2048, intermediate_size 6144, 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, sliding_window null, attention_chunk_size null, vocab_size 151936, eos_token_id 151643, and tie_word_embeddings true." }, { "label": "Qwen3 1.7B Base safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base/resolve/ea980cb0a6c2ae4b936e82123acc929f1cec04c1/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 35648-byte header with 310 BF16 tensors totaling 1720574976 parameters / 3.441149952 GB. model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 parameters / 0.622329856 GB. The checkpoint has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for the base pretrained repo, separate from the existing instruction-tuned profile." }, { "id": "qwen--qwen3-1-7b-gptq-int8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-1.7B-GPTQ-Int8", "title": "Qwen3 1.7B GPTQ Int8", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int8 Qwen3 1.7B repo.", "model_family": "qwen3-dense-gptq", "base_model_proof": { "base_model": "Qwen/Qwen3-1.7B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served GPTQ config, audited BF16 base profile comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The API metadata and model card identify Qwen/Qwen3-1.7B as the base model. Manual comparison against the audited BF16 instruction profile found matching Qwen3ForCausalLM geometry, context fields, no-sliding-attention settings, vocabulary size, and tied embedding setting; the GPTQ repo changes torch_dtype to float16 and adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-1-7b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.720574976, "swept_params_b": 1.720574976, "auxiliary_resident_params_b": 0, "resident_weight_gb": 2.066958336, "swept_weight_gb": 2.066958336, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "GPTQ logical serving parameters reconstructed from qweight tensors plus F16 model tensors in model.safetensors", "swept_parameter_scope": "ordinary text decode charges all stored tensor payload bytes because the checkpoint has tied embeddings and no separate lm_head.weight", "auxiliary_scope": "no resident-only model tensor payload for ordinary text decode; the safetensors header is file overhead and is not counted as tensor traffic", "notes": "GPTQ qweight tensors are packed I32 values; logical parameter counts treat each qweight element as four 8-bit values. qzeros, g_idx, and scales are storage/serving metadata and are charged in stored-byte traffic but excluded from logical parameter counts. The quantization config records lm_head false, tie_word_embeddings true, and the safetensors header has no lm_head.weight, so model.embed_tokens.weight is the tied output projection and remains swept for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records sliding_window null and no attention_chunk_size, so this profile charges full-context K and V streams for all 28 language layers." }, "notes": "Dense Qwen3ForCausalLM GPTQ profile using the served quantized repo config. The model card rounds context to 32768 tokens, but the pinned config records 40960 max position embeddings, matching the audited BF16 instruction profile." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1.2013183760264103, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-gptq-int8-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored GPTQ bytes from the safetensors header: packed I32 qweight/qzeros/g_idx tensors plus F16 scales, embeddings, norms, and layernorm weights. Dequantization, activation traffic, compute, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and GPTQ 8-bit quantization with group_size 128, desc_act false, symmetric quantization, true_sequential true, and lm_head false. KV cache is charged as FP16." }, "evidence": [ { "label": "Qwen3 1.7B GPTQ Int8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-1.7B-GPTQ-Int8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 382fed3fa21d1a7e837b42a500564a0c4be4f060, the API reports a public Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen3-1.7B, tags 8-bit, gptq, endpoints_compatible, and region:us, current downloads 180833, and safetensors logical parameters I32: 1409286144, F16: 311288832, total: 1720574976." }, { "label": "Qwen3 1.7B GPTQ Int8 model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8/raw/382fed3fa21d1a7e837b42a500564a0c4be4f060/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as the GPTQ-quantized 8-bit Qwen3 1.7B model, records 1.7B parameters, 1.4B non-embedding parameters, 28 layers, GQA with 16 Q heads and 8 KV heads, and GPTQ 8-bit quantization. The card rounds context length to 32768 tokens; the served config records 40960 max position embeddings." }, { "label": "Qwen3 1.7B GPTQ Int8 served config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8/raw/382fed3fa21d1a7e837b42a500564a0c4be4f060/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM with hidden_size 2048, intermediate_size 6144, 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 40960, sliding_window null, tie_word_embeddings true, torch_dtype float16, vocab_size 151936, and GPTQ quantization with bits 8, group_size 128, desc_act false, symmetric quantization, pack_dtype int32, and lm_head false." }, { "label": "Qwen3 1.7B BF16 instruction profile comparison", "url": "https://huggingface.co/Qwen/Qwen3-1.7B/raw/70d244cc86ccca08cf5af4e1e306ecf908b1ad5e/config.json", "source_type": "config", "supports": [ "base_model_proof", "logical_parameter_split", "embedding_layout" ], "notes": "Manual comparison against the audited BF16 instruction profile found matching architecture, hidden size, intermediate size, layer count, attention head count, KV head count, context, vocabulary, and tied-embedding fields. The BF16 profile records that the instruction checkpoint stores a separate lm_head.weight, while this GPTQ config sets lm_head false and the GPTQ safetensors header has no lm_head.weight." }, { "label": "Qwen3 1.7B GPTQ Int8 safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8/resolve/382fed3fa21d1a7e837b42a500564a0c4be4f060/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead", "tie_word_embeddings" ], "notes": "Direct HEAD check found linked object size 2067058456 bytes. A range-read of model.safetensors found a 100112-byte header and 898 tensors totaling 2.066958336 GB of tensor payload. Dtype bytes are I32 1.422360576 GB and F16 0.644597760 GB. Stored suffix bytes are qweight 1.409286144 GB, qzeros 0.011010048 GB, g_idx 0.002064384 GB, scales 0.022020096 GB, and F16 weight tensors 0.622577664 GB. model.embed_tokens.weight contributes 0.622329856 GB, model.norm.weight contributes 0.000004096 GB, and model.layers tensors contribute 1.444624384 GB. The header has no lm_head.weight, so the tied embedding/output matrix is swept for ordinary text decode and no model tensor payload is resident-only." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned GPTQ config, audited BF16 instruction profile comparison, linked-object HEAD check, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate by using exact GPTQ stored bytes and by treating the tied embedding as swept output-projection traffic because no separate lm_head.weight is stored." }, { "id": "qwen--qwen3-1-7b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-1.7B", "title": "Qwen3 1.7B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 1.7B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-1.7B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the served repo records eos_token_id 151645 and 40960 max position embeddings while the base config records eos_token_id 151643 and 32768 max position embeddings. This profile therefore uses the served repo config and served safetensors headers directly." }, "architecture": { "canonical_architecture_id": "qwen3-1-7b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.031739904, "swept_params_b": 1.720574976, "auxiliary_resident_params_b": 0.311164928, "resident_weight_gb": 4.063479808, "swept_weight_gb": 3.441149952, "auxiliary_resident_weight_gb": 0.622329856, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 311 BF16 tensors totaling 2031739904 stored parameters. The config marks tie_word_embeddings true, but the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config has no sliding_window or attention_chunk_size setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 1.7B model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "architecture" ], "notes": "The model metadata identifies Qwen/Qwen3-1.7B-Base as the base model and Apache-2.0 as the license. The card rounds the model as 1.7B parameters, 1.4B non-embedding parameters, 28 layers, 16 Q heads, 8 KV heads, and 32768 context length; the served config is used for the exact context value." }, { "label": "Qwen3 1.7B config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 28 layers, hidden size 2048, 16 attention heads, 8 KV heads, 128 head dimension, null sliding_window, null attention_chunk_size, and 40960 max position embeddings." }, { "label": "Qwen3 1.7B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-1.7B", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format", "base_model_proof" ], "notes": "The API records repo SHA 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e and safetensors parameters BF16: 2031739904, total: 2031739904." }, { "label": "Qwen3 1.7B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-1.7B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across both indexed shards. Stored tensors sum to 2031739904 BF16 parameters and 4.063479808 GB, matching the index total_size. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 3.441149952 GB." }, { "label": "Qwen3 1.7B Base config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different eos_token_id and max_position_embeddings values, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from served config, base config comparison, model card/API metadata, safetensors header counts, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-14b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-14B-AWQ", "title": "Qwen3 14B AWQ", "summary": "Audited memory-side bounds profile for the Qwen3 14B AWQ 4-bit checkpoint.", "model_family": "qwen3-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-14B", "relation": "quantized", "source": "Hugging Face model card metadata and served AWQ config", "config_compatible": true, "notes": "The AWQ repo card records Qwen/Qwen3-14B as the base model. The served AWQ config records the same Qwen3ForCausalLM geometry as the audited base profile, with AWQ GEMM 4-bit quantization added." }, "architecture": { "canonical_architecture_id": "qwen3-14b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7683072, "swept_params_b": 13.99039488, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 9.976576, "swept_weight_gb": 8.42075136, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus BF16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, BF16 scales, and unquantized BF16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as their unpacked 4-bit logical parameters, while qzeros and scales are storage overhead rather than model parameters." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records sliding_window null and use_sliding_window false, so this profile charges full-context K and V streams for all 40 language layers." }, "notes": "This is a dense Qwen3 causal LM profile. Thinking/non-thinking mode affects prompting and generated content, not the memory-side decoder adapter." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, BF16 scales, and unquantized BF16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen3 14B AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-14B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 31c69efc29464b6bb0aee1398b5a7b50a99340c3, the API reports a text-generation Transformers repo with base_model Qwen/Qwen3-14B, tags 4-bit and awq, safetensors logical parameters I32: 13212057600, BF16: 1556249600, total: 14768307200, and current downloads 1699790." }, { "label": "Qwen3 14B AWQ served config", "url": "https://huggingface.co/Qwen/Qwen3-14B-AWQ/raw/31c69efc29464b6bb0aee1398b5a7b50a99340c3/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen3ForCausalLM with hidden_size 5120, 40 layers, 40 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, use_sliding_window false, sliding_window null, tie_word_embeddings false, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Qwen3 14B AWQ model card", "url": "https://huggingface.co/Qwen/Qwen3-14B-AWQ/blob/31c69efc29464b6bb0aee1398b5a7b50a99340c3/README.md", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture" ], "notes": "The model card identifies this repo as Qwen3-14B with AWQ 4-bit quantization, 40 layers, GQA with 40 Q heads and 8 KV heads, and text-generation deployment through Transformers, SGLang, and vLLM." }, { "label": "Qwen3 14B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-14B-AWQ/raw/31c69efc29464b6bb0aee1398b5a7b50a99340c3/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index metadata records total_size 9989683200 bytes, but direct range-read shard headers and HTTP content lengths show 9976576000 tensor bytes plus headers. Header spans were therefore used for exact bounds. The two shards contain 1003 tensors totaling 9.976576 GB: 6.6576384 GB I32 tensors and 3.3189376 GB BF16 tensors. Stored suffix totals are qweight 6.6060288 GB, qzeros 0.0516096 GB, scales 0.2064384 GB, and BF16 weight tensors 3.1124992 GB. model.embed_tokens.weight and lm_head.weight each have shape [151936, 5120] and contribute 1.55582464 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the served AWQ config, model card, base Qwen3-14B profile comparison, safetensors index, and direct shard header byte grouping." }, "notes": "Audited from HF API metadata, served config, model card, local base-profile comparison, safetensors index metadata, and direct safetensors shard header range reads." }, { "id": "qwen--qwen3-14b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-14B-FP8", "title": "Qwen3 14B FP8", "summary": "Audited memory-side text-decode bounds profile for the FP8 Qwen3 14B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-14B", "relation": "quantized", "source": "Hugging Face model metadata, served FP8 config, audited BF16 base config/profile comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-14B as its quantized base model and preserves the audited Qwen3ForCausalLM tensor geometry: 40 layers, hidden size 5120, intermediate size 17408, 40 attention heads, 8 KV heads, 128 head dimension, 151936 vocab size, 40960 max position embeddings, no sliding window, and untied embeddings. It adds fine-grained FP8 quantization with e4m3 format, dynamic activation scheme, and 128x128 weight blocks." }, "architecture": { "canonical_architecture_id": "qwen3-14b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7691136, "swept_params_b": 13.99120128, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 16.3261696, "swept_weight_gb": 14.77034496, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, lm_head.weight, and FP8 scale-inverse tensors needed for swept weights", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 723 tensors totaling 14.769113600 stored parameters / 16.326169600 GB of direct tensor payload. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. Layer matrices are F8_E4M3 with BF16 scale-inverse tensors; embeddings, lm_head, norms, and scale-inverse tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served FP8 config has use_sliding_window false, sliding_window null, and no FP8 KV cache scheme, so the v1 profile charges full-context BF16 K and V streams for all 40 language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.1054265030502575, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and BF16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 with quantization_config fp8/e4m3 dynamic activation scheme and 128x128 weight blocks. KV cache bytes are charged as BF16 because the config does not define a quantized KV cache." }, "evidence": [ { "label": "Qwen3 14B FP8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-14B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "serving", "downloads" ], "notes": "At commit 9a283b4a5efbc09ce247e0ae5b02b744739e525a, the API reports a public non-gated Apache-2.0 text-generation repo with base_model Qwen/Qwen3-14B, base_model:quantized metadata, text-generation-inference, endpoints_compatible, fp8, region:us, 287664 downloads, and safetensors parameters BF16: 1557056000, F8_E4M3: 13212057600, total: 14769113600." }, { "label": "Qwen3 14B FP8 served config", "url": "https://huggingface.co/Qwen/Qwen3-14B-FP8/raw/9a283b4a5efbc09ce247e0ae5b02b744739e525a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, hidden_size 5120, intermediate_size 17408, 40 layers, 40 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, use_sliding_window false, sliding_window null, tie_word_embeddings false, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 14B FP8 model card", "url": "https://huggingface.co/Qwen/Qwen3-14B-FP8/blob/9a283b4a5efbc09ce247e0ae5b02b744739e525a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "architecture", "serving" ], "notes": "The card identifies this repo as the FP8 version of Qwen3-14B, records 14.8B parameters, 13.2B non-embedding parameters, 40 layers, 40 Q heads, 8 KV heads, and deployment through Transformers, SGLang, and vLLM. Its FP8 note says the quantization method is fine-grained fp8 with block size 128 and points to quantization_config." }, { "label": "Qwen3 14B BF16 base config and profile", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/40c069824f4251a91eefaf281ebe4c544efd3e18/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "Manual comparison found no differences in the audited geometry fields between the FP8 artifact config and the audited BF16 Qwen3-14B config after excluding quantization metadata. The BF16 base profile uses the same separate model.embed_tokens.weight plus lm_head.weight layout and full-context KV geometry." }, { "label": "Qwen3 14B FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-14B-FP8/raw/9a283b4a5efbc09ce247e0ae5b02b744739e525a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving", "embedding_layout" ], "notes": "Range reads of all four safetensors shard headers found 723 tensors with 16.326169600 GB of direct tensor payload: BF16 3.114112000 GB plus F8_E4M3 13.212057600 GB. The index metadata total_size is 16.339276800 GB and the linked file payload plus headers total 16.326253296 GB, so this profile follows the existing Qwen FP8 convention and uses direct tensor spans. model.embed_tokens.weight is 1.555824640 GB resident-only for ordinary decode, while model.layers.*, model.norm.weight, lm_head.weight, and FP8 scale-inverse tensors total 14.770344960 GB of swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served config, model card, audited BF16 base config/profile comparison, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-14b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-14B", "title": "Qwen3 14B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 14B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-14B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the served repo records eos_token_id 151645 and 40960 max position embeddings while the base config records eos_token_id 151643 and 32768 max position embeddings. This profile therefore uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "qwen3-14b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.7683072, "swept_params_b": 13.99039488, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 29.5366144, "swept_weight_gb": 27.98078976, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 443 BF16 tensors totaling 14768307200 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config has sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config and direct safetensors header grouping rather than deriving structure from the model name. The model card describes 32768 native context and 131072 with YaRN; this v1 profile uses the served max_position_embeddings value of 40960." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 14B model card", "url": "https://huggingface.co/Qwen/Qwen3-14B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "architecture", "max_context_tokens" ], "notes": "The card identifies Qwen/Qwen3-14B-Base as the base model, records Apache-2.0 licensing, and describes 14.8B parameters, 13.2B non-embedding parameters, 40 layers, 40 Q heads, 8 KV heads, 32768 native context, and 131072 tokens with YaRN." }, { "label": "Qwen3 14B config", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/40c069824f4251a91eefaf281ebe4c544efd3e18/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen3ForCausalLM, bfloat16, 40 layers, hidden size 5120, 40 attention heads, 8 KV heads, 128 head dimension, sliding_window null, use_sliding_window false, tie_word_embeddings false, and 40960 max position embeddings." }, { "label": "Qwen3 14B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-14B", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format", "repo", "license", "pipeline" ], "notes": "At commit 40c069824f4251a91eefaf281ebe4c544efd3e18, the API safetensors block records BF16: 14768307200 and total: 14768307200, which this profile stores as 14.7683072B resident parameters." }, { "label": "Qwen3 14B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/40c069824f4251a91eefaf281ebe4c544efd3e18/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_size 29536614400 bytes across eight shards. Range-read shard headers found 443 BF16 tensors totaling 14768307200 parameters / 29.5366144 GB, matching the index total. model.embed_tokens.weight has shape [151936, 5120] and contributes 1.55582464 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 27.98078976 GB." }, { "label": "Qwen3 14B Base config", "url": "https://huggingface.co/Qwen/Qwen3-14B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but different eos_token_id and max_position_embeddings values, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, model card/API metadata, direct safetensors index, and range-read shard header byte grouping." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses exact stored checkpoint bytes and explicit KV geometry from the served config." }, { "id": "qwen--qwen3-235b-a22b-instruct-2507-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8", "title": "Qwen3 235B A22B Instruct 2507 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3 235B A22B Instruct 2507 MoE repo.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-235B-A22B-Instruct-2507", "relation": "quantized", "source": "Model metadata declares the BF16 repo as the quantized base, and direct config comparison found matching architecture fields.", "config_compatible": true, "notes": "Manual comparison against Qwen/Qwen3-235B-A22B-Instruct-2507 at commit ac9c66cc9b46af7306746a9250f23d47083d689e found matching memory-relevant fields: Qwen3MoeForCausalLM, qwen3_moe, 94 layers, hidden size 4096, 64 Q heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, moe_intermediate_size 1536, decoder_sparse_step 1, tie_word_embeddings false, 262144 max positions, sliding_window null, use_sliding_window false, and vocab size 151936. The FP8 repo adds quantization_config while preserving the BF16 geometry used for bounds." }, "architecture": { "canonical_architecture_id": "qwen3-235b-a22b-instruct-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 236.416915456, "main_resident_weight_gb": 235.172255744, "auxiliary_resident_weight_gb": 1.244659712, "fixed_weight_gb": 8.04813824, "routed_expert_weight_gb": 1.774407168, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The config does not record a shared expert. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived stored bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down projection weights plus BF16 scale-inverse tensors; routed_expert_weight_gb is the exact routed tensor byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 94, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so Bounds Engine v1 charges full-context BF16 K and V streams for all 94 layers." }, "notes": "Qwen3MoeForCausalLM is a text-only MoE model. This profile models ordinary autoregressive text decode through the served FP8 checkpoint and does not infer structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0055677028244416, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-fp8-moe-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 quantization with 128x128 blocks and use with transformers, SGLang, and vLLM as the original BF16 model. The config does not record a KV-cache quantization scheme, so this profile keeps full-context KV cache as BF16." }, "evidence": [ { "label": "Qwen3 235B A22B Instruct 2507 FP8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "weight_format", "base_model_proof" ], "notes": "At commit e156cb4efae43fbee1a1ab073f946a1377e6b969, the live API records a public non-gated Apache-2.0 text-generation repo with qwen3_moe, fp8, endpoints_compatible, deploy:azure, and region:us tags; base_model:quantized:Qwen/Qwen3-235B-A22B-Instruct-2507; 221064 downloads; and safetensors parameters BF16 1309010944, F8_E4M3 233798893568, total 235107904512." }, { "label": "Qwen3 235B A22B Instruct 2507 FP8 model card", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8/raw/e156cb4efae43fbee1a1ab073f946a1377e6b969/README.md", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned card states that this repo contains the FP8 version of Qwen3-235B-A22B-Instruct-2507, with 235B total parameters, 22B activated parameters, 94 layers, 64 Q heads, 4 KV heads, 128 experts, 8 activated experts, 262144 native context, SGLang/vLLM deployment examples at 262144 context, and fine-grained FP8 quantization with block size 128." }, { "label": "Qwen3 235B A22B Instruct 2507 FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8/raw/e156cb4efae43fbee1a1ab073f946a1377e6b969/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, tie_word_embeddings false, 94 layers, hidden_size 4096, 64 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, moe_intermediate_size 1536, norm_topk_prob true, output_router_logits false, sliding_window null, use_sliding_window false, max_position_embeddings 262144, and FP8 quantization with dynamic activations, e4m3 weights, and 128x128 weight blocks." }, { "label": "Qwen3 235B A22B Instruct 2507 BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507/raw/ac9c66cc9b46af7306746a9250f23d47083d689e/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences between the FP8 and BF16 repos for the audited memory-relevant fields: architectures, model_type, runtime dtype, tie_word_embeddings, layer count, hidden size, intermediate sizes, attention heads, KV heads, head_dim, expert count, experts per token, decoder_sparse_step, max_position_embeddings, sliding_window, use_sliding_window, norm_topk_prob, output_router_logits, and vocab size." }, { "label": "Qwen3 235B A22B Instruct 2507 FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8/raw/e156cb4efae43fbee1a1ab073f946a1377e6b969/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 236416915456 bytes across 24 shards. Range-read shard headers found 73417 tensors totaling 236.416915456 GB stored bytes and 235107904512 logical parameters: BF16 2.618021888 GB and F8_E4M3 233.798893568 GB. model.embed_tokens.weight has shape [151936, 4096] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 227.124117504 GB, exactly 1.774407168 GB per routed expert across all 94 layers. Non-expert fixed decode tensors, including attention, router/gate, norms, and lm_head.weight, sum to 8.048138240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned model card, pinned served FP8 config, pinned BF16 base config comparison, safetensors index, and direct range-read safetensors shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. On 128GB local hardware it is expected to return resident_not_fit, but larger-memory hardware can now produce profile-backed FP8 MoE bounds without falling back to rounded metadata estimates." }, { "id": "qwen--qwen3-235b-a22b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-235B-A22B", "title": "Qwen3 235B A22B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3 235B A22B MoE repo.", "model_family": "qwen3-moe", "architecture": { "canonical_architecture_id": "qwen3-235b-a22b", "max_context_tokens": 40960, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 470.18726912, "main_resident_weight_gb": 468.942609408, "auxiliary_resident_weight_gb": 1.244659712, "fixed_weight_gb": 14.749817856, "routed_expert_weight_gb": 3.548381184, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "shared_expert_notes": "The config does not record a shared expert. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Expert tensors are stored as separate per-expert gate, up, and down projection matrices; routed_expert_weight_gb is the exact routed tensor byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 94, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all 94 layers." }, "notes": "Qwen3MoeForCausalLM is a text-only MoE model. This profile models ordinary autoregressive text decode through the served BF16 checkpoint." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The API safetensors block and range-read shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 235B A22B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-235B-A22B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "total_params_b", "routed_experts", "routed_experts_per_token", "revision" ], "notes": "At commit 8efa61729e24bd65b1d152b5ab5409052aa80e65, the live API records a public Apache-2.0 text-generation repo with qwen3_moe, endpoints_compatible, eval-results, and region:us tags; 916836 downloads; and safetensors parameters BF16: 235093634560." }, { "label": "Qwen3 235B A22B config", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B/raw/8efa61729e24bd65b1d152b5ab5409052aa80e65/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16, tie_word_embeddings false, 94 layers, hidden_size 4096, 64 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, moe_intermediate_size 1536, sliding_window null, use_sliding_window false, and max_position_embeddings 40960." }, { "label": "Qwen3 235B A22B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B/raw/8efa61729e24bd65b1d152b5ab5409052aa80e65/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 470187269120 bytes across 118 shards. Range-read shard headers found 36945 BF16 tensors totaling 235093634560 parameters / 470.187269120 GB, matching the index total. model.embed_tokens.weight has shape [151936, 4096] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors are model.layers.*.mlp.experts.*.{down_proj,gate_proj,up_proj}.weight and sum to 454.192791552 GB, exactly 3.548381184 GB per routed expert across all 94 layers. Non-expert fixed decode tensors, including lm_head.weight, sum to 14.749817856 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, the served config, model card, local scrape row, and direct safetensors header range reads for all 118 shards." }, "notes": "This profile is for ordinary text decode bounds. On 128GB local hardware it is expected to return resident_not_fit, but the audited profile prevents fallback estimates for this large BF16 MoE checkpoint." }, { "id": "qwen--qwen3-30b-a3b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B-FP8", "title": "Qwen3 30B A3B FP8", "summary": "Audited memory-side bounds profile for the official FP8 Qwen3 30B A3B MoE repo.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B", "relation": "quantized", "source": "Model card FP8 checkpoint statement, base_model metadata, served FP8 config, and direct comparison with the audited BF16 profile", "config_compatible": true, "notes": "Manual comparison found matching core architecture fields between this FP8 repo and the BF16 Qwen3-30B-A3B repo; the FP8 artifact adds quantization_config while preserving the served model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b", "max_context_tokens": 40960, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.444792832, "main_resident_weight_gb": 31.20013312, "auxiliary_resident_weight_gb": 1.244659712, "fixed_weight_gb": 2.202025984, "routed_expert_weight_gb": 0.226547712, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_f32_f8_e4m3", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived stored bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices plus F32 scale-inverse tensors; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert groups." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The FP8 config records sliding_window null and use_sliding_window false, matching the BF16 Qwen3 30B A3B config, so the v1 profile charges full-context BF16 K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM FP8 profile using the served FP8 repo config, direct BF16 profile comparison, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored F8_E4M3 matrix payloads plus F32 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 quantization with block size 128 and use with Transformers, SGLang, and vLLM as the original BF16 model. The config does not record a KV-cache quantization scheme, so this profile keeps full-context KV cache as BF16." }, "evidence": [ { "label": "Qwen3 30B A3B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At commit d206ba732169f29bb77fbf80fc2c4b81d4d30782, the API records an Apache-2.0 public text-generation repo with base_model Qwen/Qwen3-30B-A3B, region:us, and safetensors parameters F32 636948480, F8_E4M3 29896998912, and total 30533947392. Current downloads are 508657. The card states 30.5B total parameters, 3.3B activated parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, 32768 native context, 131072 context with YaRN, and fine-grained FP8 quantization with block size 128." }, { "label": "Qwen3 30B A3B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8/raw/d206ba732169f29bb77fbf80fc2c4b81d4d30782/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 6144, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 40960, rope_theta 1000000, and FP8 quantization with dynamic activation quantization and 128x128 weight blocks." }, { "label": "Qwen3 30B A3B BF16 profile comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "The existing audited BF16 profile records the same served architecture fields: Qwen3MoeForCausalLM, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 40960, and vocab_size 151936. The FP8 repo adds quantization_config and changes stored tensor dtypes/bytes." }, { "label": "Qwen3 30B A3B FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8/raw/d206ba732169f29bb77fbf80fc2c4b81d4d30782/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 32444792832 bytes across seven shards. Range-read shard headers found 37491 tensors, matching the index tensor count, and stored tensors sum exactly to 32.444792832 GB: F8_E4M3 29.896998912 GB and F32 2.547793920 GB. Linked-object HEAD checks resolved all seven shards to the pinned commit with combined linked size 32.449487768 GB and 4694936 bytes of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight is F32 with shape [151936, 2048] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same F32 shape and remains in fixed decode traffic. Expert tensors sum to 28.998107136 GB and divide exactly into 128 uniform expert groups of 0.226547712 GB. Non-expert fixed decode tensors including lm_head.weight sum to 2.202025984 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card text, served FP8 config, existing BF16 profile comparison, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MoE FP8 profile for production profile-backed bounds. It deliberately keeps KV cache BF16 because neither the config nor the model card requests FP8 KV cache." }, { "id": "qwen--qwen3-30b-a3b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B-GPTQ-Int4", "title": "Qwen3 30B A3B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int4 Qwen3 30B A3B MoE repo.", "model_family": "qwen3-moe-gptq", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B", "relation": "quantized", "source": "Hugging Face base_model metadata, model card, served GPTQ config, quantize_config.json, direct base-config comparison, and safetensors header metadata", "config_compatible": true, "notes": "The GPTQ repo records Qwen/Qwen3-30B-A3B as its quantized base model. Manual comparison against the audited BF16 base config found matching served architecture fields used by the bounds profile; the GPTQ repo changes torch_dtype to float16 and adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b", "max_context_tokens": 40960, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.924176384, "main_resident_weight_gb": 16.301846528, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.1205632, "routed_expert_weight_gb": 0.118603776, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_gptq_i32_f16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full input embedding matrix for each generated token", "notes": "Header-derived stored bytes are used because the package mixes packed GPTQ I32 tensors with F16 embeddings, output head, router, layer norm, scale, and small attention tensors. Routed experts are stored as qweight, qzeros, g_idx, and scales tensors and divide into 128 uniform expert groups." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served GPTQ config records sliding_window null and use_sliding_window false, matching the audited BF16 Qwen3 30B A3B config, so the v1 profile charges full-context FP16 K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM GPTQ profile using the served quantized repo config and direct safetensors header byte grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5543072321705084, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-gptq-int4-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored GPTQ bytes from the safetensors header: packed I32 qweight/qzeros/g_idx tensors plus F16 scales, embeddings, output head, router, norms, and attention side tensors. Dequantization, activation traffic, router compute, expert compute, cache writes, and scheduler behavior are outside this memory-side bound.", "notes": "The model card documents GPTQ 4-bit quantization and deployment with Transformers, SGLang, and vLLM. Neither the config nor quantize_config.json records a quantized KV-cache scheme, so this profile charges full-context KV cache as FP16." }, "evidence": [ { "label": "Qwen3 30B A3B GPTQ Int4 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 9b534e4318b7ebc3c961a839f13eb18b1833f441, the API reports a public Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen3-30B-A3B, base_model:quantized metadata, tags 4-bit, gptq, endpoints_compatible, region:us, current downloads 136678, and safetensors logical parameters I32 29896998912, F16 635123712, total 30532122624." }, { "label": "Qwen3 30B A3B GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "architecture", "context_notes", "serving" ], "notes": "The model card identifies the repo as GPTQ 4-bit Qwen3-30B-A3B and records 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, GQA with 32 Q heads and 4 KV heads, 128 experts, 8 activated experts, 32768 native context, optional 131072-token YaRN context, and SGLang/vLLM deployment examples." }, { "label": "Qwen3 30B A3B GPTQ Int4 served config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/raw/9b534e4318b7ebc3c961a839f13eb18b1833f441/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen3MoeForCausalLM with model_type qwen3_moe, hidden_size 2048, intermediate_size 6144, moe_intermediate_size 768, 48 layers, 32 attention heads, 4 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, 128 experts, 8 experts per token, decoder_sparse_step 1, tie_word_embeddings false, vocab_size 151936, torch_dtype float16, and GPTQ quantization with bits 4, group_size 128, desc_act false, symmetric quantization, true_sequential true, and checkpoint_format gptq." }, { "label": "Qwen3 30B A3B GPTQ Int4 quantize config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/raw/9b534e4318b7ebc3c961a839f13eb18b1833f441/quantize_config.json", "source_type": "config", "supports": [ "serving", "quantization" ], "notes": "quantize_config.json matches the config quantization block: bits 4, group_size 128, damp_percent 0.01, desc_act false, static_groups false, symmetric quantization, true_sequential true, quant_method gptq, and checkpoint_format gptq." }, { "label": "Qwen3 30B A3B BF16 base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison against the audited BF16 base config found matching architectures, model_type, hidden size, intermediate sizes, layer count, attention head count, KV head count, head dimension, max position embeddings, sliding-window fields, expert counts, experts per token, decoder_sparse_step, absent shared_expert_intermediate_size, vocabulary size, and tied-embedding setting." }, { "label": "Qwen3 30B A3B GPTQ Int4 safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/resolve/9b534e4318b7ebc3c961a839f13eb18b1833f441/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Direct range-read of the single model.safetensors header found a 9080000-byte header and 74739 tensors with stored tensor payload summing to 16.924176384 GB: 15.186788352 GB I32 and 1.737388032 GB F16. model.embed_tokens.weight is F16 [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is separate F16 [151936, 2048] and remains in fixed decode traffic. Routed expert tensors sum to 15.181283328 GB and divide exactly into 128 expert groups of 0.118603776 GB. Non-expert fixed decode tensors including attention, routers, layer norms, final norm, and lm_head sum to 1.120563200 GB. The linked file size is 16933256392 bytes, so safetensors header/container overhead is outside the tensor-payload resident bound." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, quantize_config.json, audited BF16 base config comparison, direct safetensors header byte grouping, and linked-object HEAD metadata." }, "notes": "This profile supersedes the scraped metadata estimate by charging exact GPTQ stored bytes and separating resident-only input embedding bytes from ordinary per-token swept text-decode traffic." }, { "id": "qwen--qwen3-30b-a3b-instruct-2507-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8", "title": "Qwen3 30B A3B Instruct 2507 FP8", "summary": "Audited memory-side bounds profile for the official FP8 Qwen3 30B A3B Instruct 2507 MoE repo.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B-Instruct-2507", "relation": "quantized", "source": "Model card FP8 checkpoint statement, base_model metadata, served FP8 config, and direct config comparison with the audited BF16 2507 profile", "config_compatible": true, "notes": "Manual comparison found matching core architecture fields between this FP8 repo and the BF16 Qwen3-30B-A3B-Instruct-2507 repo; the FP8 artifact adds quantization_config while preserving the BF16 model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.170895872, "main_resident_weight_gb": 30.548566016, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.553997824, "routed_expert_weight_gb": 0.226520064, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived stored bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices plus scale-inverse tensors; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert groups." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The FP8 config records sliding_window null and use_sliding_window false, matching the BF16 2507 config, so the v1 profile charges full-context BF16 K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM FP8 profile using the served 2507 FP8 repo config, direct BF16 config comparison, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 quantization with block size 128 and use with Transformers, SGLang, and vLLM as the original BF16 model. The config does not record a KV-cache quantization scheme, so this profile keeps full-context KV cache as BF16." }, "evidence": [ { "label": "Qwen3 30B A3B Instruct 2507 FP8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At commit 5a5a776300a41aaa681dd7ff0106608ef2bc90db, the API records an Apache-2.0 public text-generation repo with base_model Qwen/Qwen3-30B-A3B-Instruct-2507, endpoints_compatible, fp8, deploy:azure, region:us, and safetensors parameters BF16 636948480, F8_E4M3 29896998912, total 30533947392. Current downloads are 304783. The card states 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, 262144 native context, and fine-grained FP8 quantization with block size 128." }, { "label": "Qwen3 30B A3B Instruct 2507 FP8 model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8/raw/5a5a776300a41aaa681dd7ff0106608ef2bc90db/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving" ], "notes": "The card says this repo contains the FP8 version of Qwen3-30B-A3B-Instruct-2507, supports only non-thinking mode, has 262144 native context, and should be served with SGLang context-length 262144 or vLLM max-model-len 262144. It explicitly says the FP8 checkpoint uses fine-grained FP8 quantization with block size 128 and can be used with Transformers, SGLang, and vLLM as the original BF16 model." }, { "label": "Qwen3 30B A3B Instruct 2507 FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8/raw/5a5a776300a41aaa681dd7ff0106608ef2bc90db/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, intermediate_size 6144, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, and FP8 quantization with dynamic activation quantization and 128x128 weight blocks." }, { "label": "Qwen3 30B A3B Instruct 2507 BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences between the FP8 and BF16 repos for the audited architecture fields: architectures, model_type, hidden size, intermediate size, layer count, attention heads, KV heads, head_dim, expert count, experts per token, MoE intermediate size, decoder sparse step, max_position_embeddings, max_window_layers, sliding_window, use_sliding_window, tie_word_embeddings, runtime dtype, vocab_size, rope_theta, rope_scaling, norm_topk_prob, and mlp_only_layers. The FP8 repo adds quantization_config." }, { "label": "Qwen3 30B A3B Instruct 2507 FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8/raw/5a5a776300a41aaa681dd7ff0106608ef2bc90db/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 31170895872 bytes across 4 shards. Range-read shard headers found 37491 tensors, matching the index tensor count, and stored tensors sum exactly to 31.170895872 GB: F8_E4M3 29.896998912 GB and BF16 1.273896960 GB. Linked-object HEAD checks resolved all four shards to the pinned commit with combined linked size 31.175618584 GB and 0.004722712 GB of safetensors header/container overhead outside tensor payloads. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 28.994568192 GB and divide exactly into 128 uniform expert groups of 0.226520064 GB. Non-expert fixed decode tensors including lm_head.weight sum to 1.553997824 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card text, served FP8 config, direct BF16 config comparison, safetensors index, linked-object HEAD checks, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MoE FP8 profile for production profile-backed bounds. It deliberately keeps KV cache BF16 because neither the config nor the model card requests FP8 KV cache." }, { "id": "qwen--qwen3-30b-a3b-instruct-2507", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B-Instruct-2507", "title": "Qwen3 30B A3B Instruct 2507 BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 30B A3B Instruct 2507 MoE repo.", "model_family": "qwen3-moe", "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 61.064245248, "main_resident_weight_gb": 60.441915392, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 2.459856896, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served 2507 repo config and direct safetensors header grouping rather than deriving structure from the model name. The model card mentions optional up-to-1M-token serving with MInference and Dual Chunk Attention; this profile uses the native served max_position_embeddings value of 262144." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 30B A3B Instruct 2507 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Instruct-2507", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit 0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe, the public API records an Apache-2.0 text-generation repo with qwen3_moe tags, region:us, and safetensors parameters BF16: 30532122624. Current downloads were 1092820 when audited." }, { "label": "Qwen3 30B A3B Instruct 2507 model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The card describes the non-thinking 2507 model as 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, and 262144 native context. It separately describes optional ultra-long context serving up to 1M tokens with MInference and Dual Chunk Attention." }, { "label": "Qwen3 30B A3B Instruct 2507 config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, and vocab_size 151936." }, { "label": "Qwen3 30B A3B Instruct 2507 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 61064245248 bytes across 16 shards. Range-read shard headers found 18867 BF16 tensors totaling 30532122624 parameters / 61.064245248 GB, matching the index total. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 57.982058496 GB and divide exactly into 128 uniform expert groups of 0.452984832 GB. Non-expert fixed decode tensors including lm_head.weight sum to 2.459856896 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MoE BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-30b-a3b-thinking-2507", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B-Thinking-2507", "title": "Qwen3 30B A3B Thinking 2507 BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 30B A3B Thinking 2507 MoE repo.", "model_family": "qwen3-moe", "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 61.064245248, "main_resident_weight_gb": 60.441915392, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 2.459856896, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served Thinking 2507 repo config and direct safetensors header grouping rather than deriving structure from the model name. The model card describes this repo as thinking-only and documents optional up-to-1M-token serving with MInference and Dual Chunk Attention; this profile uses the native served max_position_embeddings value of 262144." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 30B A3B Thinking 2507 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Thinking-2507", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "resident_weight_gb", "weight_format" ], "notes": "At commit 144afc2f379b542fdd4e85a1fcd5e1f79112d95d, the public API records an Apache-2.0 text-generation repo with qwen3_moe tags, eval-results, endpoints_compatible, deploy:azure, region:us, and safetensors parameters BF16: 30532122624. Current downloads are 212751." }, { "label": "Qwen3 30B A3B Thinking 2507 model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving" ], "notes": "The card describes Qwen3-30B-A3B-Thinking-2507 as a thinking-only model and records 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, and 262144 native context. It separately documents optional ultra-long context serving up to about 1M tokens with MInference and Dual Chunk Attention." }, { "label": "Qwen3 30B A3B Thinking 2507 config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16, 48 layers, hidden_size 2048, intermediate_size 6144, moe_intermediate_size 768, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, norm_topk_prob true, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, vocab_size 151936, attention_bias false, and use_cache true." }, { "label": "Qwen3 30B A3B Thinking 2507 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json", "source_type": "manual_review", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter" ], "notes": "Manual comparison against the already audited Qwen/Qwen3-30B-A3B-Instruct-2507 config found no differences in checked architecture fields used by the bounds profile: model class, model type, dtype, embedding tying, hidden size, layer count, attention heads, KV heads, head dimension, expert count, experts per token, native context, RoPE theta, and sliding-window settings all match." }, { "label": "Qwen3 30B A3B Thinking 2507 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 61064245248 bytes across 16 shards. Range-read shard headers found 18867 BF16 tensors totaling 30532122624 parameters / 61.064245248 GB, matching the index total. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 57.982058496 GB and divide exactly into 128 uniform expert groups of 0.452984832 GB. Non-expert fixed decode tensors including lm_head.weight sum to 2.459856896 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served config, sibling config comparison, safetensors index, and direct range-read shard header byte grouping." }, "notes": "This is a self-contained MoE BF16 profile for production profile-backed bounds. It does not inherit from the Instruct 2507 profile, although the audited architecture and tensor-byte fields match exactly." }, { "id": "qwen--qwen3-30b-a3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-30B-A3B", "title": "Qwen3 30B A3B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 30B A3B MoE repo.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "The served Instruct and Base configs have matching tensor geometry, MoE routing, dtype, and sliding-window settings. The served repo records max_position_embeddings 40960 and eos_token_id 151645 while the base config records 32768 and 151643, so this profile uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b", "max_context_tokens": 40960, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 61.064245248, "main_resident_weight_gb": 60.441915392, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 2.459856896, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served repo config and direct safetensors header grouping rather than deriving structure from the model name. The model card describes 32768 native context and 131072 with YaRN; this v1 profile uses the served max_position_embeddings value of 40960." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 30B A3B model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "total_params_b", "active_params_b", "layers", "kv_heads", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The card records base_model Qwen/Qwen3-30B-A3B-Base, Apache-2.0 licensing, 30.5B total parameters, 3.3B activated parameters, 29.9B non-embedding parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, 32768 native context, and 131072 context with YaRN. It also notes that the default config uses max_position_embeddings 40960." }, { "label": "Qwen3 30B A3B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "The HF CLI model info response records repo SHA ad44e777bcd18fa416d9da3bd8f70d33ebb85d39, lastModified 2025-07-26T03:45:17Z, pipeline text-generation, base_model tags for Qwen/Qwen3-30B-A3B-Base, license:apache-2.0, and safetensors parameters BF16: 30532122624." }, { "label": "Qwen3 30B A3B config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, and max_position_embeddings 40960." }, { "label": "Qwen3 30B A3B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 61064245248 bytes across 16 shards. Range-read shard headers found 18867 BF16 tensors totaling 30532122624 parameters / 61.064245248 GB, matching the index total. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 57.982058496 GB and divide exactly into 128 uniform expert groups of 0.452984832 GB. Non-expert fixed decode tensors including lm_head.weight sum to 2.459856896 GB." }, { "label": "Qwen3 30B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Base/raw/1b75feb79f60b8dc6c5bc769a898c206a1c6a4f9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching tensor geometry, MoE routing, dtype, and sliding-window fields. Only max_position_embeddings and eos_token_id differ, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF CLI model info, model card, served config, base config comparison, safetensors index, range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained MoE BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-32b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-32B-AWQ", "title": "Qwen3 32B AWQ", "summary": "Audited memory-side text-decode bounds profile for the official Qwen3 32B AWQ 4-bit checkpoint.", "model_family": "qwen3-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-32B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served AWQ config, current BF16 base config comparison, and safetensors shard headers", "config_compatible": true, "notes": "The AWQ repo card/API metadata record Qwen/Qwen3-32B as the quantized base model. Manual comparison found matching checked architecture fields between the current BF16 base config and the AWQ artifact config. The AWQ artifact adds quantization_config while preserving the base Qwen3ForCausalLM text geometry." }, "architecture": { "canonical_architecture_id": "qwen3-32b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.762123264, "swept_params_b": 31.984210944, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 19.325298688, "swept_weight_gb": 17.769474048, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "logical AWQ model parameters represented by safetensors qweight plus BF16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, BF16 scales, and unquantized BF16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records sliding_window null and use_sliding_window false, so this profile charges full-context K and V streams for all 64 language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served AWQ config and direct safetensors header grouping. The model card describes 32768 native context and 131072 with YaRN; this v1 profile uses the served max_position_embeddings value of 40960." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5898671014779806, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-qwen3-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, BF16 scales, and unquantized BF16 tensors from safetensors headers. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. The safetensors headers store BF16 non-quantized tensors and scales, but exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Qwen3 32B AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-32B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit 0499c3ac83fdef8810b907a23894ba91e95eddd8, the live API reports a public text-generation Transformers repo with base_model Qwen/Qwen3-32B, Apache-2.0 license, 4-bit and awq tags, region:us, safetensors logical parameters I32: 31205621760 and BF16: 1556501504, total: 32762123264, and current downloads 1602159." }, { "label": "Qwen3 32B AWQ served config", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ/raw/0499c3ac83fdef8810b907a23894ba91e95eddd8/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records Qwen3ForCausalLM with hidden_size 5120, intermediate_size 25600, 64 layers, 64 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Qwen3 32B AWQ model card", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ", "source_type": "model_card", "supports": [ "base_model", "quantization", "architecture", "context_notes" ], "notes": "The model card identifies this repo as AWQ 4-bit Qwen3-32B, records 32.8B parameters, 31.2B non-embedding parameters, 64 layers, GQA with 64 Q heads and 8 KV heads, 32768 native context and 131072 context with YaRN, and recommends SGLang >= 0.4.6.post1 or vLLM >= 0.8.5 for deployment." }, { "label": "Qwen3 32B BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "The current BF16 base API records commit 9216db5781bf21249d130ec9da846c4624c16137 and safetensors BF16 total 32762123264 parameters. Manual comparison found matching checked architecture fields between the base config and AWQ artifact: Qwen3ForCausalLM, hidden_size 5120, intermediate_size 25600, 64 layers, 64 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, and rope_theta 1000000." }, { "label": "Qwen3 32B BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen3-32B", "source_type": "manual_review", "supports": [ "architecture_compatibility", "kv_adapter" ], "notes": "The already audited Qwen/Qwen3-32B profile records the same full-context text KV geometry used here: 64 text layers, 8 KV heads, 128 head dimension, 40960 served max positions, untied embeddings, and separate lm_head output projection." }, { "label": "Qwen3 32B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ/raw/0499c3ac83fdef8810b907a23894ba91e95eddd8/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index metadata reports total_size 19338405888 bytes across four shards, while direct range-read headers and linked-object HEAD checks show 19325298688 bytes of tensor payload. The profile uses direct header tensor spans for traffic. Headers contain 1603 tensors totaling 19.325298688 GB: 15.724707840 GB I32 tensors and 3.600590848 GB BF16 tensors. Stored suffix totals are qweight 15.602810880 GB, qzeros 0.121896960 GB, scales 0.487587840 GB, and BF16 weight tensors 3.113003008 GB. model.embed_tokens.weight and lm_head.weight each have shape [151936, 5120] and contribute 1.555824640 GB." }, { "label": "Qwen3 32B AWQ linked-object HEAD checks", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ/tree/0499c3ac83fdef8810b907a23894ba91e95eddd8", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for all four safetensors shards resolved to linked sizes 4966384688, 4998723112, 4949671152, and 4410702792 bytes. The linked file sizes include safetensors JSON header overhead; direct data_offsets provide the tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served AWQ config, model card, current BF16 base config comparison, existing BF16 profile, safetensors index metadata, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "Use this profile for the official Qwen3 32B AWQ artifact. Do not substitute a flat 0.5 byte/parameter estimate; the exact stored qzeros, scales, embedding, and lm_head tensors materially change resident and swept traffic." }, { "id": "qwen--qwen3-32b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-32B-FP8", "title": "Qwen3 32B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3 32B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-32B", "relation": "quantized", "source": "Hugging Face model metadata, served FP8 config, model card, audited BF16 base config/profile comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-32B as its quantized base model and preserves the audited Qwen3ForCausalLM tensor geometry: 64 layers, hidden size 5120, intermediate size 25600, 64 attention heads, 8 KV heads, 128 head dimension, 151936 vocab size, 40960 max position embeddings, no sliding window, and untied embeddings. It adds fine-grained FP8 quantization with e4m3 format, dynamic activation scheme, and 128x128 weight blocks." }, "architecture": { "canonical_architecture_id": "qwen3-32b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.764027904, "swept_params_b": 31.986115584, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 34.322434048, "swept_weight_gb": 32.766609408, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, lm_head.weight, and FP8 scale-inverse tensors needed for swept weights", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 1155 tensors totaling 32.764027904 stored parameters / 34.322434048 GB of direct tensor payload. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. Layer matrices are F8_E4M3 with BF16 scale-inverse tensors; embeddings, lm_head, norms, and scale-inverse tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served FP8 config has use_sliding_window false, sliding_window null, and no FP8 KV cache scheme, so the v1 profile charges full-context BF16 K and V streams for all 64 language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0475645469649273, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and BF16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 with quantization_config fp8/e4m3 dynamic activation scheme and 128x128 weight blocks. KV cache bytes are charged as BF16 because the config does not define a quantized KV cache." }, "evidence": [ { "label": "Qwen3 32B FP8 API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-32B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "serving", "downloads" ], "notes": "At commit aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df, the API reports a public non-gated Apache-2.0 text-generation repo with base_model Qwen/Qwen3-32B, base_model:quantized metadata, text-generation-inference, endpoints_compatible, fp8, region:us, live downloads 99892, and safetensors parameters BF16: 1558406144, F8_E4M3: 31205621760, total: 32764027904. The catalog row keeps the earlier over-100k scrape count because the model was already in the queued audit set and the catalog schema currently requires at least 100000 downloads." }, { "label": "Qwen3 32B FP8 served config", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8/raw/aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, hidden_size 5120, intermediate_size 25600, 64 layers, 64 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, use_sliding_window false, sliding_window null, tie_word_embeddings false, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 32B FP8 model card", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8/blob/aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "architecture", "serving" ], "notes": "The card identifies this repo as the FP8 version of Qwen3-32B, records 32.8B parameters, 31.2B non-embedding parameters, 64 layers, GQA with 64 Q heads and 8 KV heads, and deployment through Transformers, SGLang, and vLLM. Its FP8 note says the quantization method is fine-grained fp8 with block size 128 and points to quantization_config." }, { "label": "Qwen3 32B BF16 base config and profile", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "Manual comparison found no differences in the audited geometry fields between the FP8 artifact config and the audited BF16 Qwen3-32B config after excluding quantization metadata. The BF16 base profile uses the same separate model.embed_tokens.weight plus lm_head.weight layout and full-context KV geometry." }, { "label": "Qwen3 32B FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8/raw/aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving", "embedding_layout" ], "notes": "Range reads of all seven safetensors shard headers found 1155 tensors with 34.322434048 GB of direct tensor payload: BF16 3.116812288 GB plus F8_E4M3 31.205621760 GB. The index metadata total_size is 34.335541248 GB and the linked file payload plus headers total 34.322567640 GB, so this profile follows the existing Qwen FP8 convention and uses direct tensor spans. model.embed_tokens.weight is 1.555824640 GB resident-only for ordinary decode, while model.layers.*, model.norm.weight, lm_head.weight, and FP8 scale-inverse tensors total 32.766609408 GB of swept decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, served config, model card, audited BF16 base config/profile comparison, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-32b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-32B", "title": "Qwen3 32B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 32B repo.", "model_family": "qwen3-dense", "architecture": { "canonical_architecture_id": "qwen3-32b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.762123264, "swept_params_b": 31.984210944, "auxiliary_resident_params_b": 0.77791232, "resident_weight_gb": 65.524246528, "swept_weight_gb": 63.968421888, "auxiliary_resident_weight_gb": 1.55582464, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 707 BF16 tensors totaling 32762123264 stored parameters. The served config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config has sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config and direct safetensors header grouping rather than deriving structure from the model name. The model card describes 32768 native context and 131072 with YaRN; this v1 profile uses the served max_position_embeddings value of 40960." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 32B model card", "url": "https://huggingface.co/Qwen/Qwen3-32B", "source_type": "model_card", "supports": [ "repo", "license", "architecture", "max_context_tokens" ], "notes": "The card identifies an Apache-2.0 text-generation repo and rounds the model as 32.8B parameters, 31.2B non-embedding parameters, 64 layers, 64 Q heads, 8 KV heads, 32768 native context, and 131072 tokens with YaRN." }, { "label": "Qwen3 32B config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "serving" ], "notes": "The served config records Qwen3ForCausalLM, bfloat16, 64 layers, hidden size 5120, 64 attention heads, 8 KV heads, 128 head dimension, sliding_window null, use_sliding_window false, tie_word_embeddings false, and 40960 max position embeddings." }, { "label": "Qwen3 32B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-32B", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format", "repo" ], "notes": "The HF CLI model info response records repo SHA 9216db5781bf21249d130ec9da846c4624c16137, lastModified 2025-07-26T03:45:22Z, pipeline text-generation, tags including license:apache-2.0, and safetensors parameters BF16: 32762123264." }, { "label": "Qwen3 32B safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index records total_size 65524246528 bytes across 17 shards. Range-read shard headers found 707 BF16 tensors totaling 32762123264 parameters / 65.524246528 GB, matching the index total. model.embed_tokens.weight has shape [151936, 5120] and contributes 1.55582464 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 63.968421888 GB." }, { "label": "Qwen3 32B base-model access check", "url": "https://huggingface.co/Qwen/Qwen3-32B-Base", "source_type": "manual_review", "supports": [ "base_model_proof_absent" ], "notes": "The Qwen3 32B API/card metadata does not expose base_model. The HF CLI reported Qwen/Qwen3-32B-Base not found, so this profile intentionally does not include a base_model_proof object." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, direct safetensors index, and range-read shard header byte grouping." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses exact stored checkpoint bytes and explicit KV geometry from the served config." }, { "id": "qwen--qwen3-4b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-AWQ", "title": "Qwen3 4B AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen3 4B checkpoint.", "model_family": "qwen3-dense-awq", "base_model_proof": { "base_model": "Qwen/Qwen3-4B", "relation": "quantized", "source": "Hugging Face model metadata, served AWQ config, base config comparison, and direct safetensors header range read", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen3-4B as its quantized base model. Manual comparison found no differences in checked architecture fields between the AWQ config and the current base config after excluding torch_dtype, transformers_version, and quantization_config." }, "architecture": { "canonical_architecture_id": "qwen3-4b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.022468096, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0, "resident_weight_gb": 2.665925632, "swept_weight_gb": 2.665925632, "auxiliary_resident_weight_gb": 0, "resident_parameter_scope": "logical Qwen3 4B parameter count represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode charges all stored tensors because the config ties embeddings and the safetensors file has no separate lm_head.weight tensor", "auxiliary_scope": "none; model.embed_tokens.weight is the tied output projection and remains swept for ordinary decode", "notes": "Bounds use exact stored bytes from the safetensors header because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, and unquantized BF16 embeddings/norm tensors. Logical parameter counts match the BF16 base model; exact resident/swept byte fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so this profile charges full-context K and V streams for all 36 language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6627586766072886, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized BF16 tensors from the safetensors header. AWQ dequantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen3 4B AWQ API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 74d4bd2bd4bff9cafc9345221320bffb08b406a3, the API records a public non-gated Apache-2.0 text-generation Transformers repo with base_model Qwen/Qwen3-4B, 4-bit and AWQ tags, endpoints_compatible, region:us, 508651 downloads, and logical safetensors parameters I32: 3633315840, BF16: 389152256, total: 4022468096." }, { "label": "Qwen3 4B AWQ config", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ/raw/74d4bd2bd4bff9cafc9345221320bffb08b406a3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen3ForCausalLM, float16, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 36 layers, 32 attention heads, 8 KV heads, head dimension 128, hidden size 2560, intermediate size 9728, 40960 max position embeddings, tie_word_embeddings true, use_sliding_window false, and sliding_window null." }, { "label": "Qwen3 4B AWQ model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ/blob/74d4bd2bd4bff9cafc9345221320bffb08b406a3/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "quantization", "context_notes" ], "notes": "The card records base_model Qwen/Qwen3-4B, AWQ 4-bit quantization, Qwen3 4B architecture notes, SGLang and vLLM serving examples, and the default config context allocation of 40960 tokens with YaRN guidance for longer contexts." }, { "label": "Qwen3 4B base config", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/1cfa9a7208912126459214e8b04321603b3df60c/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the current Qwen/Qwen3-4B config after excluding quantization_config, torch_dtype, and transformers_version." }, { "label": "Qwen3 4B AWQ safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ/resolve/74d4bd2bd4bff9cafc9345221320bffb08b406a3/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The single safetensors file was range-read directly. The linked file size is 2.666027672 GB with a 102032-byte header and 2.665925632 GB of tensor bytes. The header records 902 tensors: 1.816657920 GB I32 qweight tensors, 0.014192640 GB I32 qzeros tensors, 0.056770560 GB F16 scales tensors, and 0.778304512 GB BF16 tensors. model.embed_tokens.weight is BF16 with shape [151936, 2560] and contributes 0.777912320 GB. model.norm.weight contributes 0.000005120 GB. The file has no lm_head.weight, and tie_word_embeddings is true, so the embedding table remains swept as the tied output projection." }, { "label": "Qwen3 4B BF16 profile", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "embedding_layout" ], "notes": "This AWQ profile uses the already audited Qwen3 4B dense text geometry and full-context KV layout because the AWQ config preserves the same architecture fields and does not enable KV-cache quantization." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served AWQ config, model card, current base config comparison, direct safetensors header byte grouping, and the existing audited Qwen3 4B BF16 profile." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the artifact as ideal 4-bit dense weights and did not account for exact AWQ qzeros, scales, and unquantized BF16 tensors." }, { "id": "qwen--qwen3-4b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-Base", "title": "Qwen3 4B Base BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 4B Base repo.", "model_family": "qwen3-dense", "architecture": { "canonical_architecture_id": "qwen3-4b-base", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense", "total_params_b": 4.022468096, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.38895616, "non_embedding_params_b": 3.633511936, "notes": "Range-read safetensors headers record 4022468096 BF16 stored parameters across three shards. model.embed_tokens.weight contributes 388956160 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served base repo config and safetensors headers directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, the HF API safetensors metadata records only BF16 parameters, and all range-read safetensors shard headers record BF16 tensors." }, "evidence": [ { "label": "Qwen3 4B Base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Base", "source_type": "model_card", "supports": [ "repo", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit 906bfd4b4dc7f14ee4320094d8b41684abff8539, the API records a public Apache-2.0 text-generation repo with qwen3, safetensors, endpoints_compatible, and region:us tags, current downloads 587668, and safetensors parameters BF16 4022468096, total 4022468096." }, { "label": "Qwen3 4B Base config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base/raw/906bfd4b4dc7f14ee4320094d8b41684abff8539/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, hidden size 2560, intermediate size 9728, 36 layers, 32 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, rope_theta 1000000, vocab size 151936, tie_word_embeddings true, use_sliding_window false, and sliding_window null." }, { "label": "Qwen3 4B Base safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base/raw/906bfd4b4dc7f14ee4320094d8b41684abff8539/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "serving" ], "notes": "The index lists three safetensors shards. Range-read shard headers record 398 BF16 tensors totaling 4022468096 parameters and 8.044936192 GB tensor payload bytes. model.embed_tokens.weight has shape [151936, 2560] and contributes 388956160 parameters / 0.77791232 GB. The header has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. First and last layer k_proj/v_proj tensors have shape [1024, 2560], confirming the 8 KV heads by 128 head dimension geometry." }, { "label": "Qwen3 4B finetune comparison", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "manual_review", "supports": [ "architecture_compatibility" ], "notes": "Manual comparison with the audited Qwen/Qwen3-4B profile found matching tensor geometry and stored parameter counts, but different context/RoPE settings. The base config records 32768 max position embeddings and rope_theta 1000000; Qwen/Qwen3-4B records 40960 max position embeddings." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, direct range-read safetensors shard headers, existing Qwen3 4B profile comparison, and local scrape row." }, "notes": "This is a self-contained dense BF16 base-model profile for production profile-backed bounds. It does not inherit the Qwen/Qwen3-4B finetune context setting." }, { "id": "qwen--qwen3-4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-GGUF", "title": "Qwen3 4B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the official Q4_K_M GGUF artifact of Qwen3 4B.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-4B", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Qwen base config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-4B. The selected Q4_K_M GGUF header records the same Qwen3 text geometry as the Qwen config, with 36 full-context text layers, tied embeddings, and no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "qwen3-4b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.022468096, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0, "resident_weight_gb": 2.497280256, "swept_weight_gb": 2.491323904, "auxiliary_resident_weight_gb": 0.005956352, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for Qwen3-4B-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The profile targets Qwen3-4B-Q4_K_M.gguf because the live HF API gguf.totalFileSize matches that linked object. Header tensor spans total 2.491323904 GB, while the linked file size is 2.497280256 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight. It has no output.weight, mmproj, vision, audio, MTP, or draft tensor." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Qwen3 config and selected GGUF metadata record full-context attention geometry with 36 layers, 8 KV heads, and 128 key/value head dimensions. The served config does not define sliding-window attention or recurrent state." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6208328310877919, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the official Q4_K_M GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters. Default GGUF KV is modeled as FP16 unless a workload profile explicitly chooses quantized KV." }, "evidence": [ { "label": "Qwen3 4B GGUF HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit bc640142c66e1fdd12af0bd68f40445458f3869b records base_model Qwen/Qwen3-4B, Apache-2.0 licensing, text-generation GGUF packaging, region:us, 388276 downloads, GGUF architecture qwen3, context length 40960, gguf.total 4022468096, and gguf.totalFileSize 2497280256. The API totalFileSize matches Qwen3-4B-Q4_K_M.gguf, so this profile targets that artifact." }, { "label": "Qwen3 4B GGUF model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model Qwen/Qwen3-4B, available quantizations q4_K_M, q5_0, q5_K_M, q6_K, and q8_0, and llama.cpp and Ollama examples that use Q8_0. Because the HF API selects Q4_K_M via totalFileSize and no normal-serving default overrides it, this profile targets Q4_K_M." }, { "label": "Qwen3 4B config", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/1cfa9a7208912126459214e8b04321603b3df60c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3ForCausalLM, bfloat16, tie_word_embeddings true, 36 layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, and no sliding-window attention." }, { "label": "Qwen3 4B BF16 profile", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "manual_review", "supports": [ "base_model_proof", "kv_adapter", "embedding_layout" ], "notes": "The already audited BF16 profile records the same geometry and 4022468096 stored parameters. It also verifies that the source safetensors package has no separate lm_head.weight and that tied token embeddings are part of ordinary output-projection traffic." }, { "label": "Qwen3 4B Q4_K_M GGUF linked-object HEAD check", "url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "The linked object advertises x-linked-size 2497280256 bytes and final content-length 2497280256 bytes, matching HF API gguf.totalFileSize for the selected artifact." }, { "label": "Qwen3 4B Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 28 metadata entries and 398 tensors. The selected file is 2.497280256 GB, with tensor payloads starting at byte 5956352. Tensor spans total 2.491323904 GB across 4022468096 logical elements: token_embd.weight 0.319065600 GB, blk.* tensors 2.172248064 GB, and output_norm.weight 0.000010240 GB. Tensor spans split into Q4_K 1.765048320 GB, Q6_K 0.725491200 GB, and F32 0.000784384 GB. Metadata/tokenizer/header/file overhead accounts for 0.005956352 GB. The header records qwen3.block_count 36, context_length 40960, embedding_length 2560, feed_forward_length 9728, attention.head_count 32, attention.head_count_kv 8, attention key/value length 128, rope.freq_base 1000000, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Qwen config, existing BF16 profile comparison, selected linked-object HEAD check, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the official main Q4_K_M GGUF text artifact. Do not infer other quantization siblings such as Q8_0 from this selected-artifact profile." }, { "id": "qwen--qwen3-4b-instruct-2507-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-Instruct-2507-FP8", "title": "Qwen3 4B Instruct 2507 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3 4B Instruct 2507 repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-4B-Instruct-2507", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-4B-Instruct-2507 as its base model. Manual comparison found no geometry differences between the FP8 config and the audited BF16 base config for architecture fields used by this profile; the FP8 repo only adds quantization_config." }, "architecture": { "canonical_architecture_id": "qwen3-4b-instruct-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.411646016, "swept_params_b": 4.022689856, "auxiliary_resident_params_b": 0.38895616, "resident_weight_gb": 5.189976192, "swept_weight_gb": 4.412063872, "auxiliary_resident_weight_gb": 0.77791232, "resident_parameter_scope": "safetensors_header_stored_fp8_f16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings true, but this FP8 safetensors artifact stores separate model.embed_tokens.weight and lm_head.weight tensors. The stored scale-inverse tensors and norms are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.1764262529625404, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads plus F16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card documents fine-grained FP8 quantization with block size 128. The config records torch_dtype bfloat16, quant_method fp8, fmt e4m3, dynamic activation quantization, 128x128 weight blocks, and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 4B Instruct 2507 FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Instruct-2507-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 8591804019c8b22094c3b5b4454e0edc05dffc98, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen3-4B-Instruct-2507, transformers/safetensors/qwen3/fp8 tags, region:us, 888237 downloads, and safetensors parameters split across F16: 778330176 and F8_E4M3: 3633315840, total 4411646016." }, { "label": "Qwen3 4B Instruct 2507 FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507-FP8/raw/8591804019c8b22094c3b5b4454e0edc05dffc98/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, 36 layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, rope_theta 5000000, vocab size 151936, tie_word_embeddings true, use_sliding_window false, sliding_window null, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 4B Instruct 2507 BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited BF16 Qwen3-4B-Instruct-2507 profile found matching served architecture geometry and the same native 262144 context. Unlike the BF16 base, this FP8 package stores an untied lm_head.weight tensor even though the config records tie_word_embeddings true." }, { "label": "Qwen3 4B Instruct 2507 FP8 safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507-FP8/resolve/8591804019c8b22094c3b5b4454e0edc05dffc98/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 77064-byte header with 651 tensors totaling 4411646016 parameters / 5.189976192 GB: 3633315840 F8_E4M3 parameters / 3.63331584 GB plus 778330176 F16 parameters / 1.556660352 GB. model.embed_tokens.weight has shape [151936, 2560] and contributes 388956160 parameters / 0.77791232 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 4022689856 parameters / 4.412063872 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, served FP8 config, base profile comparison, direct safetensors header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-4b-instruct-2507", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-Instruct-2507", "title": "Qwen3 4B Instruct 2507 BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 4B Instruct 2507 repo.", "model_family": "qwen3-dense", "architecture": { "canonical_architecture_id": "qwen3-4b-instruct-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense", "total_params_b": 4.022468096, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.38895616, "non_embedding_params_b": 3.633511936, "notes": "Safetensors headers record 4022468096 BF16 stored parameters across 398 tensors. model.embed_tokens.weight contributes 388956160 parameters; tie_word_embeddings is true and there is no separate lm_head tensor, leaving 3633511936 non-embedding parameters. The v1 dense adapter charges all resident weights as decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served 2507 repo config and safetensors headers directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, the HF API safetensors metadata records only BF16 parameters, and all range-read safetensors shard headers record BF16 tensors." }, "evidence": [ { "label": "Qwen3 4B Instruct 2507 model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family" ], "notes": "The card identifies Qwen3-4B-Instruct-2507 as an updated non-thinking Qwen3 4B instruction model, Apache-2.0 licensed, text-generation packaged, with 36 layers, 32 query heads, 8 KV heads, and 262144 native context." }, { "label": "Qwen3 4B Instruct 2507 config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 36 layers, hidden size 2560, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, rope_theta 5000000, tie_word_embeddings true, use_sliding_window false, and sliding_window null." }, { "label": "Qwen3 4B Instruct 2507 Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Instruct-2507", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit cdbee75f17c01a7cc42f958dc650907174af0554 records safetensors parameters BF16: 4022468096 and total: 4022468096." }, { "label": "Qwen3 4B Instruct 2507 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "serving" ], "notes": "The index lists three safetensors shards. Range-read shard headers record 398 BF16 tensors totaling 4022468096 parameters and 8044936192 tensor bytes. model.embed_tokens.weight has shape [151936, 2560], first and last layer k_proj/v_proj tensors have shape [1024, 2560], confirming the 8 KV heads by 128 head dimension geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, model card, HF API metadata, hf CLI-downloaded evidence files, and direct range-read safetensors shard headers." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. The HF API does not expose a base_model tag for this repo, so the profile does not inherit from the older Qwen3 4B profile." }, { "id": "qwen--qwen3-4b-thinking-2507-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-Thinking-2507-FP8", "title": "Qwen3 4B Thinking 2507 FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3 4B Thinking 2507 repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-4B-Thinking-2507", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, audited BF16 base profile, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-4B-Thinking-2507 as its base model. Manual comparison found no geometry differences between the FP8 config and the audited BF16 Thinking 2507 base profile for architecture fields used by this profile; the FP8 repo adds quantization_config and stores a separate lm_head tensor." }, "architecture": { "canonical_architecture_id": "qwen3-4b-thinking-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.411646016, "swept_params_b": 4.022689856, "auxiliary_resident_params_b": 0.38895616, "resident_weight_gb": 5.189976192, "swept_weight_gb": 4.412063872, "auxiliary_resident_weight_gb": 0.77791232, "resident_parameter_scope": "safetensors_header_stored_fp8_f16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings true, but this FP8 safetensors artifact stores separate model.embed_tokens.weight and lm_head.weight tensors. The stored scale-inverse tensors and norms are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.1764262529625404, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads plus F16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card documents fine-grained FP8 quantization with block size 128. The config records torch_dtype bfloat16, quant_method fp8, fmt e4m3, dynamic activation quantization, 128x128 weight blocks, and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 4B Thinking 2507 FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Thinking-2507-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 953532f942706930ec4bb870569932ef63038fdf, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen3-4B-Thinking-2507, transformers/safetensors/qwen3/fp8 tags, deploy:azure, region:us, 207284 downloads, and safetensors parameters split across F16: 778330176 and F8_E4M3: 3633315840, total 4411646016. The model card states this is an FP8 quantized checkpoint with fine-grained FP8 quantization and 128 block size, usable with Transformers, SGLang, and vLLM." }, { "label": "Qwen3 4B Thinking 2507 FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507-FP8/raw/953532f942706930ec4bb870569932ef63038fdf/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, 36 layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, rope_theta 5000000, vocab size 151936, tie_word_embeddings true, use_sliding_window false, sliding_window null, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 4B Thinking 2507 BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited BF16 Qwen3-4B-Thinking-2507 profile found matching served architecture geometry and the same native 262144 context. Unlike the BF16 base, this FP8 package stores an untied lm_head.weight tensor even though the config records tie_word_embeddings true." }, { "label": "Qwen3 4B Thinking 2507 FP8 safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507-FP8/resolve/953532f942706930ec4bb870569932ef63038fdf/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 77064-byte header with 651 tensors totaling 4411646016 parameters / 5.189976192 GB: 3633315840 F8_E4M3 parameters / 3.633315840 GB plus 778330176 F16 parameters / 1.556660352 GB. model.embed_tokens.weight has shape [151936, 2560] and contributes 388956160 parameters / 0.777912320 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 4022689856 parameters / 4.412063872 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, served FP8 config, audited BF16 base profile comparison, direct safetensors header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-4b-thinking-2507", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B-Thinking-2507", "title": "Qwen3 4B Thinking 2507 BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 4B Thinking 2507 repo.", "model_family": "qwen3-dense", "architecture": { "canonical_architecture_id": "qwen3-4b-thinking-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense", "total_params_b": 4.022468096, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.38895616, "non_embedding_params_b": 3.633511936, "notes": "Safetensors headers record 4022468096 BF16 stored parameters across 398 tensors. model.embed_tokens.weight contributes 388956160 parameters; tie_word_embeddings is true and there is no separate lm_head tensor, leaving 3633511936 non-embedding parameters. Because the input embedding is tied as the output projection, the v1 dense adapter charges all resident weights as decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and sliding_window null, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served Thinking 2507 repo config and safetensors headers directly." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, the HF API safetensors metadata records only BF16 parameters, and all range-read safetensors shard headers record BF16 tensors." }, "evidence": [ { "label": "Qwen3 4B Thinking 2507 model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "max_context_tokens" ], "notes": "The card identifies Qwen3-4B-Thinking-2507 as a thinking-only Qwen3 4B model, Apache-2.0 licensed, text-generation packaged, with 4.0B parameters, 3.6B non-embedding parameters, 36 layers, 32 query heads, 8 KV heads, and 262144 native context." }, { "label": "Qwen3 4B Thinking 2507 config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507/raw/768f209d9ea81521153ed38c47d515654e938aea/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "At commit 768f209d9ea81521153ed38c47d515654e938aea, the config records Qwen3ForCausalLM, bfloat16, 36 layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, rope_theta 5000000, tie_word_embeddings true, use_sliding_window false, and sliding_window null." }, { "label": "Qwen3 4B Thinking 2507 Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Thinking-2507", "source_type": "derived_calculation", "supports": [ "repo_commit", "downloads", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 768f209d9ea81521153ed38c47d515654e938aea records a public Apache-2.0 text-generation repo with region:us, 554589 downloads when audited, safetensors parameters BF16: 4022468096, and total: 4022468096." }, { "label": "Qwen3 4B Thinking 2507 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507", "source_type": "manual_review", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter" ], "notes": "Manual comparison against the already audited Qwen/Qwen3-4B-Instruct-2507 config found no differences in architecture fields used by the bounds profile: model class, dtype, embedding tying, hidden size, MLP size, layer count, attention heads, KV heads, head dimension, 262144 max context, rope theta, and sliding-window settings all match." }, { "label": "Qwen3 4B Thinking 2507 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507/raw/768f209d9ea81521153ed38c47d515654e938aea/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "non_embedding_params_b", "weight_format", "serving" ], "notes": "The index records total_size 8045591552 bytes across three safetensors shards. Range-read shard headers record 398 BF16 tensors totaling 4022468096 parameters and 8.044936192 tensor GB. model.embed_tokens.weight has shape [151936, 2560] and 388956160 parameters. There is no lm_head.weight tensor because embeddings are tied. First and last layer k_proj/v_proj tensors have shape [1024, 2560], confirming the 8 KV heads by 128 head dimension geometry." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, sibling config comparison, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It does not inherit from the non-thinking 2507 profile, although the audited architecture fields match exactly." }, { "id": "qwen--qwen3-4b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-4B", "title": "Qwen3 4B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 4B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-4B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the served repo records 40960 max position embeddings while the base config records 32768. This profile therefore uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "qwen3-4b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense", "total_params_b": 4.022468096, "parameter_scope": "safetensors_header_stored_bf16", "embedding_params_b": 0.38895616, "non_embedding_params_b": 3.633511936, "notes": "Range-read safetensors headers record 4022468096 BF16 stored parameters across three shards. model.embed_tokens.weight contributes 388956160 parameters; tie_word_embeddings is true and there is no separate lm_head.weight tensor, so the tied embedding/output projection remains in ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 4B model card", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen3-4B-Base as the base model and Apache-2.0 as the license." }, { "label": "Qwen3 4B config", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 36 layers, 8 KV heads, 128 head dimension, and 40960 max position embeddings." }, { "label": "Qwen3 4B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format" ], "notes": "At commit 1cfa9a7208912126459214e8b04321603b3df60c, the API safetensors block records BF16: 4022468096 and total: 4022468096, which this profile stores as 4.022468096B parameters." }, { "label": "Qwen3 4B safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "total_params_b", "embedding_params_b", "weight_format", "embedding_layout" ], "notes": "The index lists three safetensors shards. Range-read shard headers record 398 BF16 tensors totaling 4022468096 parameters and 8.044936192 GB, matching index total_size. model.embed_tokens.weight has shape [151936, 2560] and contributes 388956160 parameters / 0.77791232 GB. The index has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic." }, { "label": "Qwen3 4B Base config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but a different max_position_embeddings value, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-5-0-8b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-0.8B-Base", "title": "Qwen3.5 0.8B Base BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 0.8B pre-trained base repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-0.8B-Base", "relation": "base", "source": "Hugging Face model metadata, pinned model card, pinned config, and direct safetensors header review", "config_compatible": true, "notes": "This repo is the pre-trained base model. The pinned config and safetensors header directly provide the architecture and tensor split, so no derived sibling is used as primary evidence." }, "architecture": { "canonical_architecture_id": "qwen3-5-0-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.873438784, "swept_params_b": 0.752393024, "auxiliary_resident_params_b": 0.12104576, "resident_weight_gb": 1.746882752, "swept_weight_gb": 1.504791232, "auxiliary_resident_weight_gb": 0.24209152, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes all model.language_model tensors. The config records tie_word_embeddings true and the safetensors header has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual tensors and MTP tensors. Exact GB fields account for 2592 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 24 layers with every fourth layer marked full_attention, giving 6 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes visual and MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000005935161221, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 2592 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 0.8B Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "linear_attention_state" ], "notes": "The card identifies this repo as the pre-trained-only Qwen3.5 0.8B Base artifact, Apache-2.0 licensed, image-text-to-text, and compatible with Transformers, vLLM, and SGLang. The model overview states a causal language model with vision encoder, 0.8B language parameters, 24 layers, a 6 * (3 Gated DeltaNet + 1 Gated Attention) hidden layout, 16 value heads and 16 QK heads for Gated DeltaNet, 8 query heads and 2 KV heads for gated attention, tied LM output, MTP, and 262144 native context." }, { "label": "Qwen3.5 0.8B Base Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-0.8B-Base", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68, the API records a public non-gated Apache-2.0 image-text-to-text repo with endpoints_compatible and region:us tags, 283029 downloads, and safetensors parameters BF16: 873436192, F32: 2592, total: 873438784." }, { "label": "Qwen3.5 0.8B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, BF16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 0.8B Base safetensors index and header", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "A direct safetensors header range read found 488 stored tensors summing to 873438784 params and 1.746882752 GB: 873436192 BF16 params and 2592 F32 params. model.language_model.embed_tokens.weight is 254279680 BF16 params and 0.50855936 GB. The header has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 752393024 params and 1.504791232 GB. Auxiliary resident tensors split into model.visual 100592896 params / 0.201185792 GB and MTP 20452864 params / 0.040905728 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned config, direct safetensors header range read, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds on the pre-trained base repo. It separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-0-8b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-0.8B", "title": "Qwen3.5 0.8B BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 0.8B repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-0.8B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text tensor geometry, context fields, layer pattern, tie_word_embeddings, dtype, and vision geometry between the served repo and base config." }, "architecture": { "canonical_architecture_id": "qwen3-5-0-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.873438784, "swept_params_b": 0.752393024, "auxiliary_resident_params_b": 0.12104576, "resident_weight_gb": 1.746882752, "swept_weight_gb": 1.504791232, "auxiliary_resident_weight_gb": 0.24209152, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes all model.language_model tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual tensors and MTP tensors. Exact GB fields account for 2592 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 24 layers with every fourth layer marked full_attention, giving 6 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000005935161221, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 2592 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 0.8B model card", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "linear_attention_state" ], "notes": "The metadata identifies Qwen/Qwen3.5-0.8B-Base as the base model, Apache-2.0 licensing, and image-text-to-text packaging. The model overview states a causal language model with vision encoder, 0.8B language parameters, 24 layers, a 6 * (3 Gated DeltaNet + 1 Gated Attention) hidden layout, 16 value heads and 16 QK heads for Gated DeltaNet, 8 query heads and 2 KV heads for gated attention, tied LM output, MTP, and 262144 native context." }, { "label": "Qwen3.5 0.8B config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, BF16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 0.8B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-0.8B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 2fc06364715b967f1860aea9cf38778875588b17 records safetensors parameters BF16: 873436192, F32: 2592, and total: 873438784." }, { "label": "Qwen3.5 0.8B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text architecture, vision geometry, context, dtype, and tied-embedding fields match the served repo config." }, { "label": "Qwen3.5 0.8B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across the single shard. Stored tensors sum to 873438784 params and 1.746882752 GB: 873436192 BF16 params and 2592 F32 params. model.language_model.embed_tokens.weight is 254279680 BF16 params and 0.50855936 GB. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 752393024 params and 1.504791232 GB. Auxiliary resident tensors, defined as model.visual plus top-level MTP, sum to 121045760 BF16 params and 0.24209152 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from served config, base config comparison, HF API metadata, safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-122b-a10b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-122B-A10B-FP8", "title": "Qwen3.5 122B A10B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.5 122B A10B repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, tags, served config, and direct base-config comparison", "config_compatible": true, "notes": "The FP8 repo records Qwen/Qwen3.5-122B-A10B as its base model. Manual comparison found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry between the FP8 config and the BF16 base repo config." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 127.152313312, "main_resident_weight_gb": 122.180388864, "auxiliary_resident_weight_gb": 4.971924448, "fixed_weight_gb": 6.202116096, "routed_expert_weight_gb": 0.453040128, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 tensors with BF16 side tensors and tiny F32 tensors. Routed experts are stored as separate gate_proj, up_proj, and down_proj tensors plus scale-inverse tensors; routed_expert_weight_gb is the grouped routed tensor byte count divided by 256 uniform expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with every fourth layer using full attention, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16/F32 safetensors bytes. FP8 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config documents fine-grained FP8 weight quantization with 128x128 block size and dynamic activation quantization. It does not record an FP8 KV cache scheme, so this profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 122B A10B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9, the API records an Apache-2.0 image-text-to-text FP8 artifact derived from Qwen/Qwen3.5-122B-A10B, with safetensors parameters BF16 2058286064, F8_E4M3 123035713536, F32 6912, and total 125094006512. The repo tags include qwen3_5_moe, fp8, endpoints_compatible, deploy:azure, and region:us." }, { "label": "Qwen3.5 122B A10B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/raw/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 48 text layers, layer_types with every fourth layer full_attention, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, MTP settings, and FP8 dynamic quantization with 128x128 weight blocks." }, { "label": "Qwen3.5 122B A10B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top architecture, model_type, tie_word_embeddings, text model type, 48-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry between the FP8 quantized config and BF16 base config." }, { "label": "Qwen3.5 122B A10B FP8 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/resolve/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 39 indexed shards. Stored tensors sum to the index total_size, 127.152313312 GB, across 76656 tensors: 4.116572128 GB BF16, 123.035713536 GB F8_E4M3, and 0.000027648 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 122.180388864 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 4.971924448 GB. Routed expert tensors, defined as model.language_model.layers.*.mlp.experts.* gate/up/down weights and scale-inverse tensors, sum to 115.978272768 GB and divide exactly into 256 uniform expert indexes of 0.453040128 GB. Fixed ordinary text traffic sums to 6.202116096 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, served config, base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package does not fit on 128GB local hardware once runtime overhead is included." }, { "id": "qwen--qwen3-5-122b-a10b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-122B-A10B-GPTQ-Int4", "title": "Qwen3.5 122B A10B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int4 Qwen3.5 122B A10B package.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "quantized", "source": "Hugging Face model card, served GPTQ config, audited base config comparison, and direct safetensors-header byte audit", "config_compatible": true, "notes": "The model card identifies this repo as the GPTQ 4-bit package for Qwen3.5-122B-A10B. Manual comparison found no checked differences in top-level, text, vision, MoE, attention, and DeltaNet state geometry between the audited BF16 base config and this GPTQ config; the target adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 78.844078048, "main_resident_weight_gb": 71.368459776, "auxiliary_resident_weight_gb": 7.475618272, "fixed_weight_gb": 10.769155584, "routed_expert_weight_gb": 0.236716032, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "base logical Qwen3.5 122B A10B parameters with direct GPTQ safetensors stored-byte totals", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "shared_expert_notes": "The GPTQ config's dynamic exclusions leave attention, linear-attention, shared-expert, shared-expert-gate, visual, MTP, embedding, and lm_head tensors unquantized. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Bounds use exact stored bytes from safetensors headers because the package mixes GPTQ-packed I32 tensors, F16 scale tensors, BF16 unquantized tensors, and tiny F32 tensors. Routed expert tensors are byte-uniform across all 256 expert indexes, including qweight, qzeros, g_idx, and scales side tensors. The top-level MTP tensors are resident-only for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with every fourth layer using full_attention, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The GPTQ artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6303164604806022, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-gptq-int4-qwen3.5-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ safetensors bytes: packed I32 qweight/qzeros/g_idx tensors, F16 scale tensors, unquantized BF16 tensors, and F32 state parameters from shard headers. Dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records GPTQ 4-bit quantization with group_size 128, desc_act false, symmetric quantization, true_sequential true, BF16 text dtype, and no quantized KV cache scheme. This profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 122B A10B GPTQ Int4 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 30cd92cba9707a9aba09d1e490ed4b66b78e9606, the API records a public Apache-2.0 image-text-to-text repo with Transformers, safetensors, qwen3_5_moe, endpoints_compatible, 4-bit, gptq, deploy:azure, and region:us tags. Current downloads are 252591. The API safetensors block reports BF16 9122373104, F32 6912, I32 115964116992, and total 125086497008 logical tensor elements." }, { "label": "Qwen3.5 122B A10B GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4/raw/30cd92cba9707a9aba09d1e490ed4b66b78e9606/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "architecture" ], "notes": "The card says the repository contains int4-quantized model weights and configuration files for the post-trained model, compatible with Transformers, vLLM, SGLang, and KTransformers. It lists 122B total and 10B activated parameters, 48 layers, the hybrid 12 x (3 Gated DeltaNet + 1 Gated Attention) layout, 256 experts, 8 routed plus 1 shared expert, 262144 native context, and GPTQ 4-bit quantization." }, { "label": "Qwen3.5 122B A10B GPTQ Int4 config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4/raw/30cd92cba9707a9aba09d1e490ed4b66b78e9606/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, GPTQ 4-bit weights with group_size 128, desc_act false, symmetric quantization, true_sequential true, 48 text layers, full_attention_interval 4, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, and one MTP layer. The dynamic quantization exclusions cover attention, shared experts, MTP, visual tensors, embeddings, and lm_head." }, { "label": "Qwen3.5 122B A10B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited top-level, text_config, and vision_config geometry fields between the BF16 base config and this GPTQ artifact after excluding quantization_config and Transformers bookkeeping fields." }, { "label": "Qwen3.5 122B A10B GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4/resolve/30cd92cba9707a9aba09d1e490ed4b66b78e9606/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 39 indexed shards. Stored tensor payloads exactly match index total_size 78.844078048 GB across 149309 tensors: I32 58.787364864 GB, BF16 18.244746208 GB, F16 1.811939328 GB, and F32 0.000027648 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 71.368459776 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 7.475618272 GB. Routed expert tensors sum to 60.599304192 GB and divide exactly into 256 uniform expert indexes of 0.236716032 GB. Fixed ordinary text traffic sums to 10.769155584 GB. Linked-object HEAD checks resolved all shards to 78.864326320 GB, leaving 20248272 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned GPTQ config, BF16 base config comparison, direct safetensors index resolution, range-read safetensors shard headers, linked-object HEAD checks, representative tensor dtype inspection, and the existing Transformers qwen3_5 runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit MoE weights, missed the shared expert, used all layers as full-context KV, and did not separate resident-only multimodal, MTP, and input embedding tensors from ordinary text decode traffic." }, { "id": "qwen--qwen3-5-122b-a10b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-122B-A10B", "title": "Qwen3.5 122B A10B BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 Qwen3.5 122B A10B repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "base", "source": "Hugging Face model metadata, served config, and direct safetensors header grouping", "config_compatible": true, "notes": "This profile targets the base BF16 repo directly. Manual comparison with the already audited official FP8 derivative found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry; the BF16 repo config is used as authoritative here." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 250.17300784, "main_resident_weight_gb": 242.697389568, "auxiliary_resident_weight_gb": 7.475618272, "fixed_weight_gb": 10.769155584, "routed_expert_weight_gb": 0.905969664, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes BF16 tensors with tiny F32 tensors. Routed expert tensors are fused by expert dimension in the BF16 checkpoint; routed_expert_weight_gb is the grouped routed tensor byte count divided by 256 uniform expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with every fourth layer using full attention, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-bf16-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored BF16/F32 safetensors bytes. Activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records bfloat16 dtype and does not record a quantized KV cache scheme, so this profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 122B A10B API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit dc4d348443bc740c68e2d77492492c11606384d5, the API records an Apache-2.0 image-text-to-text repo with safetensors parameters BF16 125086490096, F32 6912, and total 125086497008. The repo tags include qwen3_5_moe, endpoints_compatible, deploy:azure, and region:us. Current downloads were 784041 when audited." }, { "label": "Qwen3.5 122B A10B config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 48 text layers, layer_types with every fourth layer full_attention, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, and one MTP layer." }, { "label": "Qwen3.5 122B A10B BF16 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/resolve/dc4d348443bc740c68e2d77492492c11606384d5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 39 indexed shards. Stored tensors sum to index total_size 250.173007840 GB across 1949 tensors: BF16 250.172980192 GB and F32 0.000027648 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 242.697389568 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 7.475618272 GB. Routed expert tensors sum to 231.928233984 GB and divide into 256 uniform expert indexes of 0.905969664 GB. Fixed ordinary text traffic, including shared experts, linear attention, self-attention, norms, and lm_head, sums to 10.769155584 GB." }, { "label": "Qwen3.5 122B A10B FP8 config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/raw/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison with the audited official FP8 derivative found matching top architecture, model_type, tie_word_embeddings, text model type, 48-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served BF16 config, FP8 derivative config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package does not fit on 128GB local hardware once runtime overhead is included." }, { "id": "qwen--qwen3-5-27b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-27B-FP8", "title": "Qwen3.5 27B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.5 27B repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3.5-27B as its base model. Manual comparison found matching core text and vision geometry: 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 attention head dimension, resident vision tower, one MTP layer, and 262144 native context." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.782935472, "swept_params_b": 25.626084864, "auxiliary_resident_params_b": 2.156850608, "resident_weight_gb": 30.86668016, "swept_weight_gb": 26.925223424, "auxiliary_resident_weight_gb": 3.941456736, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. Header-derived bytes are used because the package mixes F8_E4M3 tensors with BF16 tensors and a small F32 linear-attention tensor set." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The FP8 artifact preserves the base Qwen3.5 text architecture, so quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16/F32 safetensors bytes. FP8 dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with 128 block size. Its standard serving examples do not require FP8 KV cache, and the config does not define a KV-cache quantization scheme, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "Qwen3.5 27B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-27B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "The API records repo SHA 97f5941bf617e31c5e237364a8602ce3f03a551a, Apache-2.0 licensing, image-text-to-text pipeline, base_model:Qwen/Qwen3.5-27B, and safetensors parameters split across BF16: 3083719344, F8_E4M3: 24699207680, and F32: 8448. The card states fine-grained FP8 quantization with block size 128, 27B language parameters, 262144 native context, MTP training, and compatibility with Transformers, vLLM, SGLang, and KTransformers." }, { "label": "Qwen3.5 27B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-FP8/raw/97f5941bf617e31c5e237364a8602ce3f03a551a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, FP8 e4m3 quantization, 128x128 weight block size, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo and this FP8 artifact; the FP8 artifact adds quantization_config while preserving the base architecture." }, { "label": "Qwen3.5 27B FP8 safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-FP8/raw/97f5941bf617e31c5e237364a8602ce3f03a551a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all 11 indexed files. Stored tensors sum to 27782935472 parameters and 30.866680160 GB: 24699207680 F8_E4M3 parameters, 3083719344 BF16 parameters, and 8448 F32 parameters. model.language_model.embed_tokens.weight is 1271398400 BF16 parameters / 2.5427968 GB. lm_head.weight is a separate untied BF16 tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 25626084864 parameters / 26.925223424 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 2156850608 parameters / 3.941456736 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, model card/API metadata, range-read safetensors file headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-27b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-27B-GPTQ-Int4", "title": "Qwen3.5 27B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int4 Qwen3.5 27B package.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card, served GPTQ config, and direct base-config comparison", "config_compatible": true, "notes": "The model card title and license link identify this repo as the GPTQ Int4 package for Qwen3.5-27B. Manual comparison found no checked differences in top-level, text_config, layer_types, and vision_config geometry fields between the audited BF16 base config and this GPTQ config; the target adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 30.235043808, "swept_weight_gb": 25.921388032, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "base logical Qwen3.5 27B parameters with direct GPTQ safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the package mixes GPTQ-packed I32 tensors, F16 scale tensors, unquantized BF16 tensors, and tiny F32 tensors. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes. The GPTQ dynamic exclusions leave lm_head, input embeddings, attention, MTP, and visual tensors unquantized." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The GPTQ artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 1.0883185652025984, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-gptq-int4-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ safetensors bytes: packed I32 qweight/qzeros/g_idx tensors, F16 scale tensors, unquantized BF16 tensors, and F32 state parameters from shard headers. Dequantization, activation traffic, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records GPTQ 4-bit quantization with group_size 128, desc_act false, symmetric quantization, true_sequential true, BF16 text dtype, and no quantized KV cache scheme. This profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 27B GPTQ Int4 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-27B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 8f0c09f227ae570e79617c6d9172b59df9c16081, the API records a public Apache-2.0 image-text-to-text repo with Transformers, safetensors, qwen3_5, endpoints_compatible, 4-bit, gptq, deploy:azure, and region:us tags. Current downloads are 288886. The API safetensors block reports I32 17112760320, BF16 10668659184, F32 8448, and total 27781427952 logical tensor elements." }, { "label": "Qwen3.5 27B GPTQ Int4 model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4/raw/8f0c09f227ae570e79617c6d9172b59df9c16081/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "architecture" ], "notes": "The card says the repository contains int4-quantized model weights and configuration files for the post-trained model, compatible with Transformers, vLLM, SGLang, and KTransformers. It lists the 27B dense language architecture, 64 layers, the 16 x (3 Gated DeltaNet + 1 Gated Attention) layout, 248320-token padded embedding/output dimensions, 262144 native context, MTP training, and GPTQ 4-bit quantization." }, { "label": "Qwen3.5 27B GPTQ Int4 config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4/raw/8f0c09f227ae570e79617c6d9172b59df9c16081/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, GPTQ 4-bit weights with group_size 128, desc_act false, symmetric quantization, true_sequential true, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, a resident vision config, and one MTP layer. The dynamic quantization exclusions cover lm_head, input embeddings, attention tensors, shared-expert regexes, MTP, and visual tensors." }, { "label": "Qwen3.5 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in 23 audited top-level, text_config, layer_types, and vision_config geometry fields between the BF16 base config and this GPTQ artifact after excluding quantization_config and Transformers bookkeeping fields." }, { "label": "Qwen3.5 27B GPTQ Int4 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4/resolve/8f0c09f227ae570e79617c6d9172b59df9c16081/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 11 indexed shards. Stored tensor payloads exactly match index total_size 30.235043808 GB across 1775 tensors: I32 8.630304768 GB, BF16 21.337318368 GB, F16 0.267386880 GB, and F32 0.000033792 GB. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 25.921388032 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 4.313655776 GB. Linked-object HEAD checks resolved all shards to 30.235264416 GB, leaving 220608 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the model card, pinned GPTQ config, BF16 base config comparison, direct safetensors index resolution, range-read safetensors shard headers, linked-object HEAD checks, and the existing Transformers qwen3_5 runtime implementation review." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights, did not include GPTQ side tensors or unquantized embeddings/output head/attention/MTP/visual tensors, and did not separate resident-only multimodal, MTP, and input embedding tensors from ordinary text decode traffic." }, { "id": "qwen--qwen3-5-27b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-27B", "title": "Qwen3.5 27B BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 27B repo.", "model_family": "qwen3.5-dense-multimodal", "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 55.5628728, "swept_weight_gb": 51.249217024, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors, top-level MTP tensors, and model.language_model.embed_tokens.weight. Exact GB fields account for 8448 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000608176082, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 8448 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 27B model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "linear_attention_state" ], "notes": "The card identifies Qwen3.5-27B as Apache-2.0 licensed and image-text-to-text packaged. The model overview states a causal language model with vision encoder, 27B language parameters, 64 layers, a 16 * (3 Gated DeltaNet + 1 Gated Attention) hidden layout, 48 value heads and 16 QK heads for Gated DeltaNet, 24 query heads and 4 KV heads for gated attention, 256 full-attention head dimension, MTP, 262144 native context, and extension up to 1010000 tokens." }, { "label": "Qwen3.5 27B config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.5 27B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-27B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit fc05daec18b0a78c049392ed2e771dde82bdf654 records safetensors parameters BF16: 27781419504, F32: 8448, and total: 27781427952." }, { "label": "Qwen3.5 27B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all 11 shards. Stored tensors sum to 27781427952 params and 55.5628728 GB: 27781419504 BF16 params and 8448 F32 params. model.language_model.embed_tokens.weight is 1271398400 BF16 params and 2.5427968 GB. lm_head.weight is a separate untied tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 25624600064 params and 51.249217024 GB. Auxiliary resident tensors, defined as model.visual plus top-level MTP plus model.language_model.embed_tokens.weight, sum to 2156827888 BF16 params and 4.313655776 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card, HF API metadata, range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation. No base-model inheritance claim is used because the Qwen/Qwen3.5-27B-Base repo returned 401 in this audit environment." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-2b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-2B-Base", "title": "Qwen3.5 2B Base BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 2B Base repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-2B-Base", "relation": "base", "source": "Hugging Face model metadata, pinned config, and direct safetensors header grouping", "config_compatible": true, "notes": "This profile is for the base repo itself. Manual comparison also found no differences in audited architecture fields between this base config and the existing audited Qwen/Qwen3.5-2B instruction-tuned config." }, "architecture": { "canonical_architecture_id": "qwen3-5-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.274069824, "swept_params_b": 1.881825088, "auxiliary_resident_params_b": 0.392244736, "resident_weight_gb": 4.548144832, "swept_weight_gb": 3.76365536, "auxiliary_resident_weight_gb": 0.784489472, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes all model.language_model tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors plus top-level mtp tensors. Exact GB fields account for 2592 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 24 layers with every fourth layer marked full_attention, giving 6 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000022796133807, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 2592 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 2B Base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-2B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "weight_format" ], "notes": "At commit b1485b2fa6dfa1287294f269f5fb618e03d52d7c, the API records an Apache-2.0 image-text-to-text repo with qwen3_5, endpoints_compatible, region:us, 231586 downloads, and safetensors parameters BF16: 2274067232, F32: 2592, total: 2274069824." }, { "label": "Qwen3.5 2B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base/raw/b1485b2fa6dfa1287294f269f5fb618e03d52d7c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and mtp_num_hidden_layers 1." }, { "label": "Qwen3.5 2B instruction config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json", "source_type": "config", "supports": [ "architecture_comparison" ], "notes": "Manual comparison found no differences in audited fields between the Base config and the existing audited Qwen/Qwen3.5-2B instruction-tuned config: model type, text layer count and pattern, attention heads, KV heads, head dimensions, linear-attention dimensions, MTP fields, tied embeddings, context length, dtype, and vision geometry all match." }, { "label": "Qwen3.5 2B Base safetensors index and shard header", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base/raw/b1485b2fa6dfa1287294f269f5fb618e03d52d7c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records one shard and metadata total_size 4548144832 bytes. Range-reading the shard header found 632 tensors totaling the same 4.548144832 GB: BF16 4.548134464 GB and F32 0.000010368 GB. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 3.763655360 GB. Auxiliary resident tensors total 0.784489472 GB: model.visual tensors 0.662833152 GB plus top-level mtp tensors 0.121656320 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, immutable config, instruction-tuned config comparison, direct safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-2b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-2B", "title": "Qwen3.5 2B BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 2B repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-2B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text tensor geometry, context fields, layer pattern, tie_word_embeddings, dtype, MTP fields, and vision geometry between the served repo and base config." }, "architecture": { "canonical_architecture_id": "qwen3-5-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.274069824, "swept_params_b": 1.881825088, "auxiliary_resident_params_b": 0.392244736, "resident_weight_gb": 4.548144832, "swept_weight_gb": 3.76365536, "auxiliary_resident_weight_gb": 0.784489472, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes all model.language_model tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual tensors and top-level MTP tensors. Exact GB fields account for 2592 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 24 layers with every fourth layer marked full_attention, giving 6 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000022796133807, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 2592 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 2B model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "linear_attention_state" ], "notes": "The model metadata identifies Qwen/Qwen3.5-2B-Base as the base model, Apache-2.0 licensing, and image-text-to-text packaging. The model card states a 2B causal language model with vision encoder, 24 text layers, a 6 * (3 Gated DeltaNet + 1 Gated Attention) hidden layout, 16 value heads and 16 QK heads for Gated DeltaNet, 8 query heads and 2 KV heads for gated attention, tied LM output, MTP, and 262144 native context." }, { "label": "Qwen3.5 2B config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and mtp_num_hidden_layers 1." }, { "label": "Qwen3.5 2B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-2B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline", "revision" ], "notes": "At commit 15852e8c16360a2fea060d615a32b45270f8a8fc, the API records an Apache-2.0 image-text-to-text repo with 1706576 downloads and safetensors parameters BF16: 2274067232, F32: 2592, total: 2274069824." }, { "label": "Qwen3.5 2B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text architecture, vision geometry, context, MTP, dtype, and tied-embedding fields match the served repo config." }, { "label": "Qwen3.5 2B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across the single shard. Stored tensors sum to 2274069824 params and 4.548144832 GB: 2274067232 BF16 params and 2592 F32 params. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 1881825088 params and 3.76365536 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp tensors, sum to 392244736 params and 0.784489472 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from served config, base config comparison, HF API metadata, safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-35b-a3b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-35B-A3B-FP8", "title": "Qwen3.5 35B A3B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.5 35B A3B repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The FP8 repo records Qwen/Qwen3.5-35B-A3B as its base model. Manual comparison found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry between the FP8 config and the BF16 base repo config." }, "architecture": { "canonical_architecture_id": "qwen3-5-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 37.454799072, "main_resident_weight_gb": 34.690869376, "auxiliary_resident_weight_gb": 2.763929696, "fixed_weight_gb": 2.474682496, "routed_expert_weight_gb": 0.12584448, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512 and the model card states 8 routed plus 1 shared activated experts. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 tensors with BF16 side tensors and tiny F32 tensors. Routed experts are stored as separate gate_proj, up_proj, and down_proj tensors plus scale-inverse tensors; routed_expert_weight_gb is the grouped routed tensor byte count divided by 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16/F32 safetensors bytes. FP8 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with block size 128. It does not require an FP8 KV cache in the serving examples, so this profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 35B A3B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "The API response at commit 9d1823d2dee688a6b25e77009dc727688c44936e records an Apache-2.0 image-text-to-text FP8 artifact derived from Qwen/Qwen3.5-35B-A3B, with safetensors parameters BF16 1500859120, F8_E4M3 34453061632, F32 4800, and total 35953925552. The card states fine-grained FP8 quantization with 128 block size, 35B total parameters, 3B activated parameters, a 10 x (3 x DeltaNet -> MoE, 1 x gated attention -> MoE) layout, 16 Q heads, 2 KV heads, 256k default context, 8 routed experts plus 1 shared expert, and vLLM/SGLang examples using 262144-token context." }, { "label": "Qwen3.5 35B A3B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8/raw/9d1823d2dee688a6b25e77009dc727688c44936e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 40 text layers, layer_types with every fourth layer full_attention, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, MTP settings, and FP8 dynamic quantization with 128x128 weight blocks." }, { "label": "Qwen3.5 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/59d61f3ce65a6d9863b86d2e96597125219dc754/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top architecture, model_type, tie_word_embeddings, text model type, 40-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry between the FP8 quantized config and BF16 base config." }, { "label": "Qwen3.5 35B A3B FP8 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8/resolve/9d1823d2dee688a6b25e77009dc727688c44936e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 14 indexed shards. Stored tensors sum to the index total_size, 37.454799072 GB, across 64196 tensors: 3.00171824 GB BF16, 34.453061632 GB F8_E4M3, and 0.0000192 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 34.690869376 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 2.763929696 GB. Routed expert tensors, defined as model.language_model.layers.*.mlp.experts.* gate/up/down weights and scale-inverse tensors, sum to 32.21618688 GB and divide exactly into 256 uniform expert indexes of 0.12584448 GB. Fixed ordinary text traffic sums to 2.474682496 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the model card, served config, base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-35b-a3b-gptq-int4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4", "title": "Qwen3.5 35B A3B GPTQ Int4", "summary": "Audited memory-side text-decode bounds profile for the official GPTQ Int4 Qwen3.5 35B A3B repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The GPTQ repo records Qwen/Qwen3.5-35B-A3B as its base model. Manual comparison found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry between the GPTQ config and the BF16 base repo config; the GPTQ repo only adds quantization_config metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.403162208, "main_resident_weight_gb": 20.803619456, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.879602816, "routed_expert_weight_gb": 0.06610944, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_gptq_i32_f16_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512 and the model card states 8 routed plus 1 shared activated experts. The GPTQ dynamic exclusion list leaves shared_expert tensors uncompressed; shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes packed GPTQ I32 tensors, F16 GPTQ scale tensors, BF16 excluded tensors, and tiny F32 tensors. The quantization dynamic list excludes lm_head, input embeddings, attention, shared experts, MTP, and visual tensors. Routed experts are stored as separate qweight, qzeros, g_idx, and scale tensors; routed_expert_weight_gb is the grouped routed tensor byte count divided by 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full_attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6787739917643982, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-gptq-moe-wna16-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored GPTQ/BF16/F16/F32 safetensors bytes. GPTQ dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The model card recommends vLLM/SGLang with --quantization moe_wna16 and 262144-token context. It does not require a quantized KV cache, so this profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 35B A3B GPTQ Int4 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "The API response at commit 3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b records an Apache-2.0 image-text-to-text GPTQ artifact derived from Qwen/Qwen3.5-35B-A3B, with current downloads 915845 and safetensors parameters I32 32212254720, BF16 3739563184, F32 4800, total 35951822704. The card states GPTQ 4-bit quantization, 35B total parameters, 3B activated parameters, a 10 x (3 x DeltaNet -> MoE, 1 x gated attention -> MoE) layout, 16 Q heads, 2 KV heads, 256k default context, 8 routed experts plus 1 shared expert, and vLLM/SGLang examples using --quantization moe_wna16." }, { "label": "Qwen3.5 35B A3B GPTQ Int4 config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4/raw/3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 40 text layers, layer_types with every fourth layer full_attention, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, MTP settings, and GPTQ quantization with bits 4, group_size 128, desc_act false, symmetric quantization, true_sequential true, and dynamic exclusions for lm_head, input embeddings, attention, shared experts, MTP, and visual tensors." }, { "label": "Qwen3.5 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/59d61f3ce65a6d9863b86d2e96597125219dc754/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top architecture, model_type, tie_word_embeddings, text model type, 40-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry between the GPTQ quantized config and BF16 base config." }, { "label": "Qwen3.5 35B A3B GPTQ Int4 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4/resolve/3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 14 indexed shards. Stored tensors sum to the index total_size, 24.403162208 GB, across 124611 tensors: 16.42070016 GB I32, 0.50331648 GB F16, 7.479126368 GB BF16, and 0.0000192 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 20.803619456 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Routed expert tensors, defined as model.language_model.layers.*.mlp.experts.* qweight, qzeros, g_idx, and scales tensors, sum to 16.92401664 GB and divide exactly into 256 uniform expert indexes of 0.06610944 GB. Fixed ordinary text traffic sums to 3.879602816 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, the model card, served config, base config comparison, direct safetensors header byte grouping, and the existing Transformers qwen3_5 runtime implementation review used by the sibling BF16/FP8 Qwen3.5 profiles." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-35b-a3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-35B-A3B", "title": "Qwen3.5 35B A3B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3.5 35B A3B instruction-tuned multimodal repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-35B-A3B-Base", "relation": "finetune", "source": "Hugging Face model card base_model metadata and direct config comparison", "config_compatible": true, "notes": "The repo records Qwen/Qwen3.5-35B-A3B-Base as its base model. Manual comparison found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry between the base and instruction-tuned repos." }, "architecture": { "canonical_architecture_id": "qwen3-5-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 71.903655008, "main_resident_weight_gb": 68.304112256, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.879602816, "routed_expert_weight_gb": 0.25165824, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. Routed experts are stored as grouped gate_up_proj and down_proj tensors; routed_expert_weight_gb is the grouped routed tensor byte count divided by 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-sglang-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside Bounds Engine v1.", "notes": "The HF API reports BF16 tensors plus a tiny F32 group; the text config records mamba_ssm_dtype float32. Full-attention KV cache is charged as BF16, and the linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 35B A3B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens", "serving" ], "notes": "At commit 59d61f3ce65a6d9863b86d2e96597125219dc754, the API records an Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen3.5-35B-A3B-Base and safetensors parameters BF16 35951817904, F32 4800, total 35951822704. The card states 35B total parameters, 3B activated, 10 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)), 16 Q heads and 2 KV heads for gated attention, 256k default context, and vLLM/SGLang examples using --max-model-len or --context-length 262144." }, { "label": "Qwen3.5 35B A3B config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/59d61f3ce65a6d9863b86d2e96597125219dc754/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with a qwen3_5_moe_text text config, 40 text layers, layer_types with every fourth layer full_attention, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 routed experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Qwen3.5 35B A3B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top architecture, model_type, tie_word_embeddings, text model type, 40-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry between the base and instruction-tuned configs." }, { "label": "Qwen3.5 35B A3B safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/resolve/59d61f3ce65a6d9863b86d2e96597125219dc754/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 14 indexed shards. Stored tensors sum to 71.903655008 GB across 1811 tensors: BF16 71.903635808 GB and F32 0.0000192 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 68.304112256 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Main routed grouped expert tensors sum to 64.42450944 GB, or 0.25165824 GB per expert index. Fixed ordinary text traffic sums to 3.879602816 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the model card, served config, base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-397b-a17b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-397B-A17B-FP8", "title": "Qwen3.5 397B A17B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.5 397B A17B repo.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-397B-A17B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and direct base-config comparison", "config_compatible": true, "notes": "The FP8 repo records Qwen/Qwen3.5-397B-A17B as its base model. Manual comparison found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry between the FP8 config and the BF16 base repo config." }, "architecture": { "canonical_architecture_id": "qwen3-5-397b-a17b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 406.12518128, "main_resident_weight_gb": 396.546877312, "auxiliary_resident_weight_gb": 9.578303968, "fixed_weight_gb": 9.952634752, "routed_expert_weight_gb": 0.75506688, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024 and the model card states 10 routed plus 1 shared activated experts. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 tensors with BF16 side tensors and tiny F32 tensors. Routed experts are stored as separate gate_proj, up_proj, and down_proj tensors plus scale-inverse tensors; routed_expert_weight_gb is the grouped routed tensor byte count divided by 512 uniform expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 15, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 60 layers with every fourth layer using full attention, giving 15 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.19316736, "read_gb_per_output_token": 0.19316736, "state_formula": "45 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16/F32 safetensors bytes. FP8 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with block size 128. The config records dynamic activation quantization with 128x128 weight blocks and does not record an FP8 KV cache scheme, so this profile charges full-attention KV as BF16 and separately charges DeltaNet recurrent state." }, "evidence": [ { "label": "Qwen3.5 397B A17B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-397B-A17B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit ea5b4f81096f3901c91dea97f81324302495781d, the API records an Apache-2.0 image-text-to-text FP8 artifact derived from Qwen/Qwen3.5-397B-A17B, with current downloads 829107, region:us, and safetensors parameters BF16 2702768432, F8_E4M3 400719609856, F32 8640, and total 403422386928. The card states fine-grained FP8 quantization with block size 128, 397B total parameters, 17B activated parameters, a 15 x (3 x DeltaNet -> MoE, 1 x gated attention -> MoE) layout, 32 Q heads, 2 KV heads, 262144 native context, and 10 routed plus 1 shared activated experts." }, { "label": "Qwen3.5 397B A17B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8/raw/ea5b4f81096f3901c91dea97f81324302495781d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 60 text layers, layer_types with every fourth layer full_attention, 15 full-attention layers, 45 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 512 experts, 10 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, MTP settings, and FP8 dynamic quantization with 128x128 weight blocks." }, { "label": "Qwen3.5 397B A17B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/8472618112abcbd45acbcdc58436aff4233c23f7/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching top architecture, model_type, tie_word_embeddings, text model type, 60-layer hybrid layout, hidden size, expert sizes, expert counts, attention head geometry, 262144 max position embeddings, DeltaNet state geometry, and vision geometry between the FP8 quantized config and BF16 base config. The only audited serving difference is the FP8 quantization_config on the quantized repo." }, { "label": "Qwen3.5 397B A17B FP8 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8/resolve/ea5b4f81096f3901c91dea97f81324302495781d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 94 indexed shards. Stored tensors sum to the index total_size, 406.12518128 GB, across 189042 tensors: 5.405536864 GB BF16, 400.719609856 GB F8_E4M3, and 0.00003456 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 396.546877312 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 9.578303968 GB. Routed expert tensors, defined as model.language_model.layers.*.mlp.experts.* gate/up/down weights and scale-inverse tensors, sum to 386.59424256 GB and divide exactly into 512 uniform expert indexes of 0.75506688 GB. Fixed ordinary text traffic sums to 9.952634752 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, served config, base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package does not fit on 128GB local hardware once runtime overhead is included." }, { "id": "qwen--qwen3-5-397b-a17b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-397B-A17B", "title": "Qwen3.5 397B A17B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3.5 397B A17B base repo.", "model_family": "qwen3.5-moe-multimodal", "architecture": { "canonical_architecture_id": "qwen3-5-397b-a17b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 806.795875168, "main_resident_weight_gb": 790.658480512, "auxiliary_resident_weight_gb": 16.137394656, "fixed_weight_gb": 17.564367232, "routed_expert_weight_gb": 1.50994944, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024 and the model card states 10 routed plus 1 shared activated experts. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the base package stores BF16 tensors plus tiny F32 recurrent scalars. Routed experts are stored as stacked per-layer BF16 tensors with shape [512, ...]; routed_expert_weight_gb is the grouped ordinary-language routed tensor byte count divided by 512 uniform expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 15, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 60 layers with every fourth layer using full attention, giving 15 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.19316736, "read_gb_per_output_token": 0.19316736, "state_formula": "45 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Bounds Engine v1 charges stored BF16/F32 safetensors bytes. Activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The config records bfloat16 text dtype and mamba_ssm_dtype float32. Full-attention KV is charged as BF16, and DeltaNet recurrent state is charged separately." }, "evidence": [ { "label": "Qwen3.5 397B A17B model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-397B-A17B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At repo SHA 8472618112abcbd45acbcdc58436aff4233c23f7, the API records a public Apache-2.0 image-text-to-text Transformers safetensors repo with qwen3_5_moe, eval-results, endpoints_compatible, and region:us tags. Current downloads are 420092. The API safetensors block reports BF16 403397920304, F32 8640, and total 403397928944. The model card states 397B total parameters, 17B activated parameters, a 15 x (3 x DeltaNet -> MoE, 1 x gated attention -> MoE) layout, 32 Q heads, 2 KV heads, 262144 native context, and 10 routed plus 1 shared activated experts." }, { "label": "Qwen3.5 397B A17B config", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/8472618112abcbd45acbcdc58436aff4233c23f7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, bfloat16 text dtype, 60 text layers, layer_types with every fourth layer full_attention, 15 full-attention layers, 45 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 512 experts, 10 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Qwen3.5 397B A17B safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/resolve/8472618112abcbd45acbcdc58436aff4233c23f7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 94 indexed shards. Stored tensors sum to the index total_size, 806.795875168 GB, across 2924 tensors: 806.795840608 GB BF16 and 0.000034560 GB F32. The dtype-derived parameter total is 403397928944, matching the API total. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 790.658480512 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 16.137394656 GB. Top-level MTP tensors account for 13.191136256 GB, visual tensors 0.912020960 GB, and the input embedding 2.034237440 GB. Ordinary-language routed expert tensors sum to 773.094113280 GB and divide into 512 uniform expert indexes of 1.509949440 GB. Fixed ordinary text traffic sums to 17.564367232 GB." }, { "label": "Qwen3.5 397B A17B FP8 config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8/raw/ea5b4f81096f3901c91dea97f81324302495781d/config.json", "source_type": "config", "supports": [ "architecture" ], "notes": "The audited FP8 profile has matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry. The FP8 repo changes serving quantization only; this BF16 base profile uses the pinned base config and base safetensors headers directly." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served config, direct safetensors header byte grouping, comparison against the existing FP8/NVFP4 derivative profiles, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package does not fit on 128GB local hardware once runtime overhead is included." }, { "id": "qwen--qwen3-5-4b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-4B-Base", "title": "Qwen3.5 4B Base BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 4B Base repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-4B-Base", "relation": "base", "source": "Hugging Face model metadata, pinned config, direct safetensors header grouping, and instruction-tuned config comparison", "config_compatible": true, "notes": "This profile is for the base repo itself. Manual comparison also found no differences across 25 audited architecture fields or the layer-type sequence between this base config and the existing audited Qwen/Qwen3.5-4B instruction-tuned config." }, "architecture": { "canonical_architecture_id": "qwen3-5-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.659865088, "swept_params_b": 4.205751296, "auxiliary_resident_params_b": 0.454113792, "resident_weight_gb": 9.319737856, "swept_weight_gb": 8.411510272, "auxiliary_resident_weight_gb": 0.908227584, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "model.visual tensors and top-level mtp tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes all model.language_model tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors plus top-level mtp tensors. Exact GB fields account for 3840 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000016481163843, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 3840 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 4B Base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-4B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "total_params_b", "weight_format" ], "notes": "At commit 1001bb4d826a52d1f399e183466143f4da7b741b, the API records an Apache-2.0 image-text-to-text repo with qwen3_5, endpoints_compatible, region:us, 197752 downloads, and safetensors parameters BF16: 4659861248, F32: 3840, total: 4659865088." }, { "label": "Qwen3.5 4B Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "linear_attention_state", "max_context_tokens" ], "notes": "The card describes the package as the pre-trained-only Qwen3.5 4B base model. It records a causal language model with vision encoder, 4B language model, 2560 hidden size, 32 layers, an 8 x (3 Gated DeltaNet layers plus 1 Gated Attention layer) layout, 16 attention heads, 4 KV heads, 256 attention head dimension, 128-dimensional DeltaNet heads, tied language-model output, MTP training, and 262144 native context." }, { "label": "Qwen3.5 4B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base/raw/1001bb4d826a52d1f399e183466143f4da7b741b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and mtp_num_hidden_layers 1." }, { "label": "Qwen3.5 4B instruction config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json", "source_type": "config", "supports": [ "architecture_comparison" ], "notes": "Manual comparison found no differences across 25 checked top-level, text, and vision architecture fields, and no difference in the layer_types sequence, between this Base config and the existing audited Qwen/Qwen3.5-4B instruction-tuned config." }, { "label": "Qwen3.5 4B Base safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base/raw/1001bb4d826a52d1f399e183466143f4da7b741b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records two shards and metadata total_size 9319737856 bytes. HEAD checks found linked shard sizes 5329398712 and 3990429344 bytes, leaving 90200 bytes of safetensors header/container overhead outside tensor payloads. Range-reading both shard headers found 738 tensors totaling the same 9.319737856 GB: BF16 9.319722496 GB and F32 0.000015360 GB. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 8.411510272 GB. Auxiliary resident tensors total 0.908227584 GB: model.visual tensors 0.667028480 GB plus top-level mtp tensors 0.241199104 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, immutable config, instruction-tuned config comparison, direct safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-4b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-4B", "title": "Qwen3.5 4B BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 4B repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-4B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text tensor geometry, context fields, layer pattern, and dtype between the served repo and base config." }, "architecture": { "canonical_architecture_id": "qwen3-5-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.659865088, "swept_params_b": 4.205751296, "auxiliary_resident_params_b": 0.454113792, "resident_weight_gb": 9.319737856, "swept_weight_gb": 8.411510272, "auxiliary_resident_weight_gb": 0.908227584, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model_safetensors_headers", "auxiliary_scope": "visual tower and MTP tensors are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors. The config records tie_word_embeddings true and the safetensors index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual tensors and MTP tensors. Exact GB fields account for 3840 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000016481163843, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 3840 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 4B model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The model metadata identifies Qwen/Qwen3.5-4B-Base as the base model, Apache-2.0 licensing, and image-text-to-text packaging." }, { "label": "Qwen3.5 4B config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Qwen3.5 4B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-4B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a records safetensors parameters BF16: 4659861248, F32: 3840, and total: 4659865088." }, { "label": "Qwen3.5 4B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text architecture and context fields match the served repo config." }, { "label": "Qwen3.5 4B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across both shards. Stored tensors sum to 4659865088 params and 9.319737856 GB. model.language_model.embed_tokens.weight is 635699200 BF16 params and 1.2713984 GB. The index has no separate lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied output embedding, sum to 4205751296 params and 8.411510272 GB. Auxiliary resident tensors, defined as visual plus MTP tensors, sum to 454113792 params and 0.908227584 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from served config, base config comparison, HF API metadata, safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP weights from per-token swept language weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-9b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-9B-Base", "title": "Qwen3.5 9B Base BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 9B Base repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B-Base", "relation": "base", "source": "Hugging Face model metadata, model card, served config, safetensors index, and direct shard-header review", "config_compatible": true, "notes": "This repo is the canonical Qwen3.5 9B base package. Manual comparison against Qwen/Qwen3.5-9B at commit c202236235762e1c871ad0ccb60c8ee5ba337b9a found matching checked architecture fields and the same safetensors parameter/byte split, but this profile uses the Base repo's own pinned config and shard headers as evidence." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 19.306216416, "swept_weight_gb": 15.873376768, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "visual tower, MTP tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset is model.language_model.embed_tokens.weight, model.visual tensors, and MTP tensors. Exact GB fields account for 3840 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000795598981, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records bfloat16 dtype and mamba_ssm_dtype float32; safetensors headers record BF16 plus 3840 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 9B Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "linear_attention_state" ], "notes": "The model card identifies a causal language model with vision encoder, Apache-2.0 licensing, 9B language parameters, hidden size 4096, padded 248320 token embedding, 32 layers, a repeated Gated DeltaNet/Gated Attention hidden layout, 32 V and 16 QK linear-attention heads at 128 dimensions, 16 Q and 4 KV full-attention heads at 256 dimensions, MTP training, and 262144 native context." }, { "label": "Qwen3.5 9B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The pinned config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, resident vision config, and MTP tensors." }, { "label": "Qwen3.5 9B Base Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-9B-Base", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 68c46c4b3498877f3ef123c856ecfde50c39f404 records downloads 198528, public non-gated access, image-text-to-text packaging, Apache-2.0 licensing, region:us, and safetensors parameters BF16: 9653100528 and F32: 3840." }, { "label": "Qwen3.5 9B Base safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all four shards. Stored tensors sum to 9653104368 params and 19.306216416 GB. model.language_model.embed_tokens.weight is 1017118720 BF16 params and 2.034237440 GB. lm_head.weight is a separate untied tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 7936684544 params and 15.873376768 GB. Auxiliary resident tensors are model.visual plus MTP plus model.language_model.embed_tokens.weight, summing to 1716419824 params and 3.432839648 GB." }, { "label": "Qwen3.5 9B instruction config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the Base and instruction configs match on checked memory-relevant fields: architecture, model types, 32 text layers, full_attention_interval 4, hidden size 4096, intermediate size 12288, 16 attention heads, 4 KV heads, 256 full-attention head dimension, linear-attention state geometry, 262144 max positions, vision config presence, and untied embeddings." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned Base model card/config, direct safetensors shard-header parameter/byte split, instruction sibling config comparison, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-5-9b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.5-9B", "title": "Qwen3.5 9B BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 Qwen3.5 9B repo.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": true, "notes": "Manual comparison found matching text tensor geometry, context fields, layer pattern, and dtype between the served repo and base config." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 19.306216416, "swept_weight_gb": 15.873376768, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "visual tower, MTP tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes visual tensors, MTP tensors, and model.language_model.embed_tokens.weight. Exact GB fields account for 3840 F32 language parameters." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000795598981, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and swept GB fields account for the small F32 tensor set.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32; safetensors headers record BF16 plus 3840 F32 parameters. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.5 9B model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "linear_attention_state" ], "notes": "The model metadata identifies Qwen/Qwen3.5-9B-Base as the base model, Apache-2.0 licensing, and image-text-to-text packaging. The model card states a 32-layer hidden layout with 8 groups of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 32 V heads and 16 QK heads for Gated DeltaNet, 128-dimensional DeltaNet heads, MTP, and 262144 native context." }, { "label": "Qwen3.5 9B config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Qwen3.5 9B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-9B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit c202236235762e1c871ad0ccb60c8ee5ba337b9a records safetensors parameters BF16: 9653100528, F32: 3840, and total: 9653104368." }, { "label": "Qwen3.5 9B Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found the relevant text architecture and context fields match the served repo config." }, { "label": "Qwen3.5 9B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all four shards. Stored tensors sum to 9653104368 params and 19.306216416 GB. model.language_model.embed_tokens.weight is 1017118720 BF16 params and 2.03423744 GB. lm_head.weight is a separate untied tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 7936684544 params and 15.873376768 GB. Auxiliary resident tensors, defined as visual plus MTP plus model.language_model.embed_tokens.weight, sum to 1716419824 params and 3.432839648 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from served config, base config comparison, HF API metadata, safetensors header parameter/byte split, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-6-27b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.6-27B-FP8", "title": "Qwen3.6 27B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.6 27B repo.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3.6-27B as its base model. Manual comparison found matching core text and vision geometry: 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 attention head dimension, resident vision tower, one MTP layer, and 262144 native context." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.782935472, "swept_params_b": 25.626084864, "auxiliary_resident_params_b": 2.156850608, "resident_weight_gb": 30.866663264, "swept_weight_gb": 26.925206528, "auxiliary_resident_weight_gb": 3.941456736, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. Header-derived bytes are used because the package mixes F8_E4M3 tensors with BF16 tensors and BF16 scale-inverse side tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The FP8 artifact preserves the base Qwen3.6 text architecture, so quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with 128 block size. Its standard vLLM and SGLang examples do not require FP8 KV cache, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "Qwen3.6 27B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.6-27B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "The API records repo SHA e89b16ebf1988b3d6befa7de50abc2d76f26eb09, Apache-2.0 licensing, image-text-to-text pipeline, base_model:Qwen/Qwen3.6-27B, and safetensors parameters split across BF16: 3083727792 and F8_E4M3: 24699207680. The card states fine-grained FP8 quantization with block size 128, 27B parameters, 262144 native context, MTP training, and text-only vLLM serving with --language-model-only." }, { "label": "Qwen3.6 27B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B-FP8/raw/e89b16ebf1988b3d6befa7de50abc2d76f26eb09/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, FP8 e4m3 quantization, 128x128 weight block size, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the base BF16 repo and this FP8 artifact; the FP8 artifact adds quantization_config while preserving the base architecture." }, { "label": "Qwen3.6 27B FP8 safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.6-27B-FP8/raw/e89b16ebf1988b3d6befa7de50abc2d76f26eb09/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all 66 indexed files. Stored tensors sum to 27782935472 parameters and 30.866663264 GB: 24699207680 F8_E4M3 parameters and 3083727792 BF16 parameters. model.language_model.embed_tokens.weight is 1271398400 BF16 parameters / 2.5427968 GB. lm_head.weight is a separate untied BF16 tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 25626084864 parameters / 26.925206528 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 2156850608 parameters / 3.941456736 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, base config comparison, model card/API metadata, range-read safetensors file headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-6-27b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.6-27B", "title": "Qwen3.6 27B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3.6 27B repo.", "model_family": "qwen3.6-dense-multimodal", "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 55.562855904, "swept_weight_gb": 51.249200128, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text config records dtype bfloat16 and mamba_ssm_dtype float32. Safetensors headers record only BF16 stored model tensors; the linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.6 27B model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family" ], "notes": "The card identifies Qwen3.6-27B as Apache-2.0 licensed, image-text-to-text packaged, with 64 layers, a 16 * (3 Gated DeltaNet + 1 Gated Attention) hidden layout, 24 query heads and 4 KV heads for gated attention, MTP, 262144 native context, and text-only serving mode that skips the vision encoder and multimodal profiling." }, { "label": "Qwen3.6 27B config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Qwen3.6 27B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.6-27B", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "The API response at commit 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 records safetensors parameters BF16: 27781427952 and total: 27781427952." }, { "label": "Qwen3.6 27B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb" ], "notes": "Safetensors headers were range-read across all 15 shards. Stored tensors sum to 27781427952 BF16 params and 55.562855904 GB. model.language_model.embed_tokens.weight is 1271398400 BF16 params and 2.5427968 GB. lm_head.weight is a separate untied tensor of the same size. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 25624600064 BF16 params and 51.249200128 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 2156827888 BF16 params and 4.313655776 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, model card, HF API metadata, range-read safetensors shard headers, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-6-35b-a3b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.6-35B-A3B-FP8", "title": "Qwen3.6 35B A3B FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3.6 35B A3B repo.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3.6-35B-A3B as its base model and keeps the same Qwen3.6 text architecture: 40 text layers, 256 routed experts, 8 experts per token, 1 shared expert, and every fourth layer using full attention." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 37.454789472, "main_resident_weight_gb": 34.690859776, "auxiliary_resident_weight_gb": 2.763929696, "fixed_weight_gb": 2.474672896, "routed_expert_weight_gb": 0.12584448, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes F8_E4M3 weights with BF16 side tensors. Routed expert tensors are byte-uniform across all 256 expert indexes, including FP8 weights and BF16 scale-inverse tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The FP8 artifact preserves the base Qwen3.6 text architecture, so quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with 128 block size. It does not require FP8 KV cache in the serving examples, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "Qwen3.6 35B A3B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.6-35B-A3B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "The API records the repo as an Apache-2.0 image-text-to-text artifact derived from Qwen/Qwen3.6-35B-A3B, with BF16 and F8_E4M3 safetensors. The card states fine-grained FP8 quantization with 128 block size, 35B total parameters, 3B activated parameters, 16 Q heads, 2 KV heads, and MTP training." }, { "label": "Qwen3.6 35B A3B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, FP8 quantization, a 40-layer text config, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, and one MTP layer." }, { "label": "Qwen3.6 35B A3B FP8 safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 42 indexed shards. Stored tensors sum to 37.454789472 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 34.690859776 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 2.763929696 GB. Main routed expert tensors sum to 32.21618688 GB, or 0.12584448 GB per expert index. Fixed ordinary text traffic sums to 2.474672896 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from model card/API metadata, served config, safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. The vLLM text-only serving command in the model card is consistent with excluding vision profiling from per-token decode traffic; FP8 weight quantization does not change the Qwen3.6 DeltaNet runtime state geometry." }, { "id": "qwen--qwen3-6-35b-a3b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3.6-35B-A3B", "title": "Qwen3.6 35B A3B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3.6 35B A3B base serving artifact.", "model_family": "qwen3.6-moe", "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 71.903645408, "main_resident_weight_gb": 68.304102656, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.25165824, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Routed expert tensors are stored as grouped expert matrices; routed_expert_weight_gb is the total routed expert tensor byte count divided by 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The language stack repeats 3 linear-attention layers followed by 1 gated-attention layer across 40 layers, giving 10 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "BF16 profile for the base Hugging Face repo. Use the NVIDIA NVFP4 repo for the worked-example local-bound row." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the base repo.", "notes": "The config dtype is bfloat16, and safetensors headers record only BF16 stored model tensors. The linear-attention recurrent runtime state is separately charged as F32 in the KV/state adapter." }, "evidence": [ { "label": "Qwen3.6 35B A3B model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "source_type": "model_card", "supports": [ "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "linear_attention_state", "max_context_tokens" ], "notes": "The overview states 35B total, 3B activated, 40 layers, 10 x (3 x Gated DeltaNet -> 1 x Gated Attention), 32 V heads and 16 QK heads for Gated DeltaNet, 128-dimensional DeltaNet heads, 16 Q heads and 2 KV heads for Gated Attention, 256 experts, 8 routed plus 1 shared expert, and 262144 native context." }, { "label": "Qwen3.6 35B A3B config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, dtype bfloat16, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, and one MTP layer." }, { "label": "Qwen3.6 35B A3B safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 26 indexed shards. Stored tensors sum to 71.903645408 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 68.304102656 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Main routed expert grouped tensors sum to 64.42450944 GB, or 0.25165824 GB per expert index. Fixed ordinary text traffic sums to 3.879593216 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from model card/API metadata, served config, range-read safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-8b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-8B-AWQ", "title": "Qwen3 8B AWQ 4-bit", "summary": "Audited memory-side bounds profile for the official AWQ 4-bit Qwen3 8B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B", "relation": "quantized", "source": "Hugging Face model metadata, served AWQ config, base config/profile comparison, and safetensors header review", "config_compatible": true, "notes": "The AWQ repo records Qwen/Qwen3-8B as its quantized base model. Manual comparison found matching Qwen3ForCausalLM tensor geometry, context, tokenizer, and attention fields between the AWQ config and the audited BF16 base config after excluding quantization metadata, torch_dtype, _name_or_path, and transformers_version." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 6.098479104, "swept_weight_gb": 4.853819392, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "logical Qwen3 8B parameters represented by the AWQ safetensors package", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and the separately stored lm_head.weight tensor", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "Bounds use exact direct tensor bytes from the safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, BF16 scales/norms, and unquantized BF16 embedding/head tensors. Exact resident and swept GB fields drive the memory-side calculation." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and no quantized KV cache scheme. The 128 head dimension is derived from hidden_size 4096 divided by 32 attention heads." }, "notes": "Dense Qwen3ForCausalLM profile using the served AWQ repo config and exact stored AWQ tensor bytes." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.744558191170762, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, BF16 scales/norms, and unquantized BF16 embedding/head tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128, zero_point true, and modules_to_not_convert null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Qwen3 8B AWQ model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format" ], "notes": "At commit 4da05a8edb55c6046cce958586c33b61da07bb79, the API records an Apache-2.0 text-generation repo with base_model Qwen/Qwen3-8B, text-generation-inference, endpoints_compatible, 4-bit, AWQ, region:us, and 378299 downloads. The API safetensors block reports logical parameters split across I32: 6945767424 and BF16: 1244967936, total 8190735360." }, { "label": "Qwen3 8B AWQ config", "url": "https://huggingface.co/Qwen/Qwen3-8B-AWQ/raw/4da05a8edb55c6046cce958586c33b61da07bb79/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format" ], "notes": "The config records Qwen3ForCausalLM, float16 runtime dtype, AWQ 4-bit GEMM quantization, group_size 128, zero_point true, modules_to_not_convert null, 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, tie_word_embeddings false, and use_sliding_window false." }, { "label": "Qwen3 8B BF16 base config and profile", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/b968826d9c46dd6066d109eabc6255188de91218/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout" ], "notes": "Manual comparison found no differences in audited geometry and context fields between the AWQ config and the current audited BF16 Qwen3-8B config after excluding quantization_config, torch_dtype, _name_or_path, and transformers_version. The BF16 base profile uses the same separate model.embed_tokens.weight plus lm_head.weight layout." }, { "label": "Qwen3 8B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-8B-AWQ/raw/4da05a8edb55c6046cce958586c33b61da07bb79/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format" ], "notes": "The index records total_size 6100576256 bytes across two shards. Range-read safetensors headers found 903 tensors totaling 6.098479104 GB of direct tensor payload: 3.500015616 GB I32 tensors and 2.598463488 GB BF16 tensors. model.embed_tokens.weight is BF16 with shape [151936, 4096] and contributes 622329856 logical parameters / 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 4.853819392 GB swept traffic. The index total_size is larger than direct tensor spans by 0.002097152 GB, so this profile uses direct shard-header tensor spans." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served AWQ config, audited BF16 base config/profile comparison, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as ideal 4-bit dense weights and undercounted stored AWQ scales, zeros, duplicate head storage, and unquantized embedding/head tensors." }, { "id": "qwen--qwen3-8b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-8B-Base", "title": "Qwen3 8B Base BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3 8B Base repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B-Base", "relation": "base", "source": "Hugging Face model metadata, served config, and direct safetensors header evidence", "config_compatible": true, "notes": "This repo is the canonical Qwen3 8B base checkpoint. Manual comparison with the already audited Qwen/Qwen3-8B fine-tune found matching core tensor geometry but a different EOS token and context length: this base config records 32768 max position embeddings while the fine-tune records 40960." }, "architecture": { "canonical_architecture_id": "qwen3-8b-base", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 16.38147072, "swept_weight_gb": 15.136811008, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and the separately stored lm_head.weight tensor", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 399 BF16 tensors totaling 8190735360 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false and no quantized KV cache scheme. The 128 head dimension is derived from hidden_size 4096 divided by 32 attention heads." }, "notes": "Dense Qwen3ForCausalLM base profile using the served base repo config rather than inheriting the fine-tuned context setting." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, the HF API safetensors metadata records only BF16 parameters, and direct safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 8B Base model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-Base", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "total_params_b", "weight_format" ], "notes": "At repo SHA 49e3418fbbbca6ecbdf9608b4d22e5a407081db4, the API records a public Apache-2.0 text-generation repo with transformers, safetensors, qwen3, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 428867. The API safetensors block reports BF16 8190735360 parameters, total 8190735360." }, { "label": "Qwen3 8B Base config", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/raw/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, tie_word_embeddings false, hidden size 4096, intermediate size 12288, 36 layers, 32 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, rope_theta 1000000, vocab size 151936, no sliding_window, and use_sliding_window false." }, { "label": "Qwen3 8B Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base", "source_type": "model_card", "supports": [ "architecture", "max_context_tokens", "non_embedding_params_b" ], "notes": "The model overview states this is the pretraining-stage Qwen3-8B-Base causal language model with 8.2B total parameters, 6.95B non-embedding parameters, 36 layers, 32 query heads, 8 KV heads, and 32768-token context length." }, { "label": "Qwen3 8B Base safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/resolve/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "The index maps 399 tensors across five shards and records total_size 16.383567872 GB. Direct range-read safetensors headers sum to 16.381470720 GB of BF16 tensor payload; the 0.002097152 GB difference is shard container/header overhead outside tensor spans. model.embed_tokens.weight is BF16 [151936, 4096] and contributes 0.622329856B parameters / 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept output-projection traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 7.568405504B parameters / 15.136811008 GB." }, { "label": "Qwen3 8B fine-tune comparison", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/b968826d9c46dd6066d109eabc6255188de91218/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited Qwen/Qwen3-8B profile found matching hidden size, intermediate size, layer count, attention head count, KV head count, head dimension, tied-embedding setting, RoPE theta, vocab size, and no-sliding-window setting. The fine-tune records EOS token 151645 and 40960 max positions, while this base repo records EOS token 151643 and 32768 max positions." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, safetensors index metadata, direct safetensors shard header range reads, and comparison against the existing Qwen3 8B fine-tune profile." }, "notes": "This is a self-contained dense BF16 base-model profile for production profile-backed bounds. It does not inherit the Qwen/Qwen3-8B fine-tune context setting." }, { "id": "qwen--qwen3-8b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-8B-FP8", "title": "Qwen3 8B FP8", "summary": "Audited memory-side text-decode bounds profile for the FP8 Qwen3 8B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B", "relation": "quantized", "source": "Hugging Face model metadata, served FP8 config, base config comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-8B as its quantized base model and preserves the served Qwen3ForCausalLM tensor geometry: 36 layers, 32 attention heads, 8 KV heads, 128 head dimension, 151936 vocab size, 40960 max position embeddings, and untied embeddings. It adds fine-grained FP8 quantization with e4m3 format and 128x128 weight blocks." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.191159296, "swept_params_b": 7.56882944, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 9.436551168, "swept_weight_gb": 8.191891456, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers.*, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 651 tensors totaling 8191159296 stored parameters / 9.436551168 GB of direct tensor payload. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. Layer matrices are mostly F8_E4M3 with BF16 scale-inverse tensors; embeddings, lm_head, and the final norm are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has use_sliding_window false and no FP8 KV cache scheme, so the v1 profile charges full-context BF16 K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served FP8 repo config and direct safetensors header byte grouping." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.1520409782053884, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and BF16 embeddings, lm_head, norms, and scale-inverse tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 with quantization_config fp8/e4m3 dynamic activation scheme and 128x128 weight blocks. KV cache bytes are charged as BF16 because the config does not define a quantized KV cache." }, "evidence": [ { "label": "Qwen3 8B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 220b46e3b2180893580a4454f21f22d3ebb187d3, the API reports an Apache-2.0 public text-generation repo with base_model Qwen/Qwen3-8B, region:us, and safetensors parameters BF16: 1245391872, F8_E4M3: 6945767424, total: 8191159296. Current downloads were 609936 when audited." }, { "label": "Qwen3 8B FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-8B-FP8/raw/220b46e3b2180893580a4454f21f22d3ebb187d3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Qwen3ForCausalLM, bfloat16 runtime dtype, 36 layers, hidden_size 4096, intermediate_size 12288, 32 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, use_sliding_window false, tie_word_embeddings false, and quantization_config quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128]." }, { "label": "Qwen3 8B BF16 base config and profile", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/b968826d9c46dd6066d109eabc6255188de91218/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout" ], "notes": "Manual comparison found no differences in the audited geometry fields between the FP8 artifact config and the current audited BF16 Qwen3-8B config after excluding quantization metadata and descriptive fields. The BF16 base profile uses the same separate model.embed_tokens.weight plus lm_head.weight layout." }, { "label": "Transformers Qwen3 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.51.0/src/transformers/models/qwen3/modeling_qwen3.py", "source_type": "manual_review", "supports": [ "kv_adapter", "embedding_layout" ], "notes": "Manual implementation review found k_proj and v_proj sized by num_key_value_heads times head_dim, cache updates for key/value states, model.embed_tokens for input lookup, and a separate lm_head projection." }, { "label": "Qwen3 8B FP8 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-8B-FP8/resolve/220b46e3b2180893580a4454f21f22d3ebb187d3/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "serving", "embedding_layout" ], "notes": "Range reads of both safetensors shard headers found 651 tensors with data_offsets exactly matching dtype times shape. Direct tensor payload is 9.436551168 GB: BF16 2.490783744 GB plus F8_E4M3 6.945767424 GB. model.embed_tokens.weight is 1.244659712 GB resident-only for ordinary decode, while model.layers.*, model.norm.weight, and lm_head.weight total 8.191891456 GB of swept decode traffic. The index metadata total_size is larger than direct tensor spans by 2097152 bytes, so this profile uses the direct shard-header tensor spans." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, BF16 base config/profile comparison, Transformers Qwen3 implementation review, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for production profile-backed bounds. It intentionally does not assume FP8 KV cache or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-8b", "version": "1.0.1", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-8B", "title": "Qwen3 8B BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3 8B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-8B-Base", "relation": "finetune", "source": "Hugging Face model metadata and base config comparison", "config_compatible": false, "notes": "Core tensor geometry matches the base config, but the served repo records 40960 max position embeddings while the base config records 32768. This profile therefore uses the served repo config directly." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 16.38147072, "swept_weight_gb": 15.136811008, "auxiliary_resident_weight_gb": 1.244659712, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 399 BF16 tensors totaling 8190735360 stored parameters. The config marks tie_word_embeddings false and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM profile using the served repo config rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 8B model card", "url": "https://huggingface.co/Qwen/Qwen3-8B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license" ], "notes": "The model metadata identifies Qwen/Qwen3-8B-Base as the base model and Apache-2.0 as the license." }, { "label": "Qwen3 8B config", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/main/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, 36 layers, 8 KV heads, 128 head dimension, and 40960 max position embeddings." }, { "label": "Qwen3 8B Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B", "source_type": "derived_calculation", "supports": [ "resident_params_b", "weight_format" ], "notes": "At commit b968826d9c46dd6066d109eabc6255188de91218, the API safetensors block records BF16: 8190735360 and total: 8190735360, which this profile stores as 8.19073536B resident parameters." }, { "label": "Qwen3 8B safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/main/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The index lists five safetensors shards. Range-read shard headers record 399 BF16 tensors totaling 8190735360 parameters and 16.38147072 GB, matching index total_size. model.embed_tokens.weight has shape [151936, 4096] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 15.136811008 GB." }, { "label": "Qwen3 8B Base config", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core tensor geometry but a different max_position_embeddings value, so this profile does not copy the base config wholesale." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-01", "notes": "Audited from the served config, base config comparison, HF API safetensors metadata, direct safetensors header grouping, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-coder-30b-a3b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8", "title": "Qwen3-Coder 30B A3B Instruct FP8", "summary": "Audited memory-side bounds profile for the official FP8 Qwen3-Coder 30B A3B Instruct MoE repo.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Model card FP8 checkpoint statement and direct config comparison with the BF16 repo", "config_compatible": true, "notes": "Manual comparison found matching core architecture fields between this FP8 repo and the BF16 Qwen3-Coder 30B A3B Instruct repo; the FP8 artifact adds quantization_config while preserving the BF16 model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 31.170895872, "main_resident_weight_gb": 30.548566016, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 1.553997824, "routed_expert_weight_gb": 0.226520064, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived stored bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices plus scale-inverse tensors; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert groups." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The FP8 config records sliding_window null and use_sliding_window false, matching the BF16 Coder config, so the v1 profile charges full-context BF16 K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM FP8 profile using the served Coder FP8 repo config, direct BF16 config comparison, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 quantization with block size 128 and use with transformers, SGLang, and vLLM as the original BF16 model. The config does not record a KV-cache quantization scheme, so this profile keeps full-context KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Coder 30B A3B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At commit dcaee4d4dfc5ee71ad501f01f530e5652438fde0, the API records an Apache-2.0 text-generation repo with qwen3_moe and fp8 tags plus safetensors parameters BF16 636948480, F8_E4M3 29896998912, and total 30533947392. Current downloads were 1,358,940 during audit. The card states 30.5B total parameters, 3.3B activated parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, 262144 native context, and fine-grained FP8 quantization with block size 128." }, { "label": "Qwen3-Coder 30B A3B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8/raw/dcaee4d4dfc5ee71ad501f01f530e5652438fde0/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, and FP8 quantization with dynamic activation quantization and 128x128 weight blocks." }, { "label": "Qwen3-Coder 30B A3B BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences between the FP8 and BF16 repos for the audited architecture fields: architectures, model_type, hidden size, intermediate size, layer count, attention heads, KV heads, head_dim, expert count, experts per token, MoE intermediate size, decoder sparse step, max_position_embeddings, max_window_layers, sliding_window, use_sliding_window, tie_word_embeddings, runtime dtype, vocab_size, rope_theta, rope_scaling, and mlp_only_layers. The FP8 repo adds quantization_config." }, { "label": "Qwen3-Coder 30B A3B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8/resolve/dcaee4d4dfc5ee71ad501f01f530e5652438fde0/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 31170895872 bytes across 4 shards. Range-read shard headers found 37491 tensors, matching the index tensor count, and stored tensors sum exactly to 31.170895872 GB: F8_E4M3 29.896998912 GB and BF16 1.27389696 GB. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 28.994568192 GB and divide exactly into 128 uniform expert groups of 0.226520064 GB. Non-expert fixed decode tensors including lm_head.weight sum to 1.553997824 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card text, served FP8 config, direct BF16 config comparison, safetensors index, range-read shard header byte grouping, and the existing audited BF16 Coder profile." }, "notes": "This is a self-contained MoE FP8 profile for production profile-backed bounds. It deliberately keeps KV cache BF16 because neither the config nor the model card request FP8 KV cache." }, { "id": "qwen--qwen3-coder-30b-a3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "title": "Qwen3-Coder 30B A3B Instruct BF16", "summary": "Audited memory-side bounds profile for the BF16 Qwen3-Coder 30B A3B Instruct MoE repo.", "model_family": "qwen3-coder-moe", "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 61.064245248, "main_resident_weight_gb": 60.441915392, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 2.459856896, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices; routed_expert_weight_gb is the total routed expert byte count divided by 128 uniform expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context K and V streams for all layers." }, "notes": "Qwen3MoeForCausalLM profile using the served Coder repo config and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3-Coder 30B A3B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-30B-A3B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit b2cff646eb4bb1d68355c01b18ae02e7cf42d120, the API records an Apache-2.0 text-generation repo with qwen3_moe tags and safetensors parameters BF16: 30532122624." }, { "label": "Qwen3-Coder 30B A3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, and rope_theta 10000000." }, { "label": "Qwen3-Coder 30B A3B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 61064245248 bytes across 16 shards. Range-read shard headers found 18867 BF16 tensors totaling 30532122624 parameters / 61.064245248 GB, matching the index total. model.embed_tokens.weight has shape [151936, 2048] and contributes 0.622329856 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 57.982058496 GB and divide exactly into 128 uniform expert groups of 0.452984832 GB. Non-expert fixed decode tensors including lm_head.weight sum to 2.459856896 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, served config, safetensors index, range-read shard header byte grouping, and local scrape row." }, "notes": "This is a self-contained MoE BF16 profile for production profile-backed bounds." }, { "id": "qwen--qwen3-coder-480b-a35b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8", "title": "Qwen3-Coder 480B A35B Instruct FP8", "summary": "Audited memory-side bounds profile for the official FP8 Qwen3-Coder 480B A35B Instruct MoE repo.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-480B-A35B-Instruct", "relation": "quantized", "source": "Model card FP8 checkpoint statement and direct config comparison with the BF16 repo", "config_compatible": true, "notes": "Manual comparison found matching audited architecture fields between this FP8 repo and the BF16 Qwen3-Coder 480B A35B Instruct repo except max_window_layers. That field is inert for this profile because both configs record use_sliding_window false and sliding_window null, so full-context KV is charged for all layers. The FP8 artifact adds quantization_config while preserving the BF16 model geometry used for bounds." }, "architecture": { "canonical_architecture_id": "qwen3-coder-480b-a35b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 482.141974528, "main_resident_weight_gb": 480.27498496, "auxiliary_resident_weight_gb": 1.866989568, "fixed_weight_gb": 12.13351936, "routed_expert_weight_gb": 2.92588416, "routed_experts": 160, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through all non-embedding tensors, with expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Header-derived stored bytes are used instead of rounded model-card parameters. Expert tensors are stored as per-expert gate/up/down matrices plus BF16 scale-inverse tensors; routed_expert_weight_gb is the total routed expert byte count divided by 160 uniform expert groups." }, "kv_adapter": { "kind": "full_context", "layers": 62, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The FP8 config records sliding_window null and use_sliding_window false, so the v1 profile charges full-context BF16 K and V streams for all 62 layers." }, "notes": "Qwen3MoeForCausalLM FP8 profile using the served Coder FP8 repo config, direct BF16 config comparison, and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-fp8-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and cache writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 quantization with block size 128 and use with transformers, SGLang, and vLLM as the original BF16 model. The config does not record a KV-cache quantization scheme, so this profile keeps full-context KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Coder 480B A35B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "At repo SHA 003f183a92fbe5b9a8325aaa8b2ae797c91dd90f, the API records an Apache-2.0 text-generation repo with qwen3_moe, fp8, endpoints_compatible, deploy:azure, and region:us tags plus safetensors parameters BF16 1957910528, F8_E4M3 478226153472, and total 480184064000. Current downloads were 531878 when audited. The card states 480B total parameters, 35B activated parameters, 62 layers, 96 Q heads, 8 KV heads, 160 routed experts, 8 activated experts, 262144 native context, and fine-grained FP8 quantization with block size 128." }, { "label": "Qwen3-Coder 480B A35B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8/raw/003f183a92fbe5b9a8325aaa8b2ae797c91dd90f/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The served config records Qwen3MoeForCausalLM, qwen3_moe, bfloat16 runtime dtype, 62 layers, hidden_size 6144, 96 attention heads, 8 KV heads, head_dim 128, 160 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, rope_theta 10000000, and FP8 quantization with dynamic activation quantization and 128x128 weight blocks." }, { "label": "Qwen3-Coder 480B A35B BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct/raw/9d90cf8fca1bf7b7acca42d3fc9ae694a2194069/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found no differences between the FP8 and BF16 repos for the audited architecture fields: architectures, model_type, hidden size, intermediate size, layer count, attention heads, KV heads, head_dim, expert count, experts per token, MoE intermediate size, decoder sparse step, max_position_embeddings, sliding_window, use_sliding_window, tie_word_embeddings, runtime dtype, vocab_size, rope_theta, rope_scaling, and mlp_only_layers. max_window_layers differs, but both configs disable sliding-window attention, so it does not change the v1 KV adapter." }, { "label": "Qwen3-Coder 480B A35B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8/resolve/003f183a92fbe5b9a8325aaa8b2ae797c91dd90f/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "The index records total_size 482141974528 bytes across 49 shards. Range-read shard headers found 60329 tensors, matching the index tensor count, and stored tensors sum exactly to 482.141974528 GB: F8_E4M3 478.226153472 GB and BF16 3.915821056 GB. model.embed_tokens.weight has shape [151936, 6144] and contributes 1.866989568 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in fixed decode traffic. Expert tensors sum to 468.141465600 GB and divide exactly into 160 uniform expert groups of 2.925884160 GB. Non-expert fixed decode tensors including lm_head.weight sum to 12.133519360 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card text, served FP8 config, direct BF16 config comparison, safetensors index, and range-read shard header byte grouping." }, "notes": "This is a self-contained MoE FP8 profile for production profile-backed bounds. It deliberately keeps KV cache BF16 because neither the config nor the model card request FP8 KV cache." }, { "id": "qwen--qwen3-coder-next-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Coder-Next-FP8", "title": "Qwen3-Coder-Next FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-Coder-Next repo.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model card statement and direct config comparison", "config_compatible": true, "notes": "The model card identifies this repository as the FP8-quantized Qwen3-Coder-Next checkpoint. Manual comparison found matching core architecture fields between the FP8 repo, Qwen/Qwen3-Coder-Next, and Qwen/Qwen3-Coder-Next-Base; the FP8 artifact only adds quantization_config." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 80.362904064, "main_resident_weight_gb": 79.740574208, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 2.421725696, "routed_expert_weight_gb": 0.151013376, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding", "auxiliary_scope": "model.embed_tokens.weight is resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 80B/3B model-card parameters. Routed expert tensors are byte-uniform across all 512 expert indexes, including FP8 weights and BF16 scale-inverse tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config generates a layer_types pattern from full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative or tool-calling behavior outside Bounds Engine v1." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-sglang-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, and state writes are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with block size 128. The config has no KV-cache quantization scheme and the vLLM/SGLang examples do not request FP8 KV cache, so this profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Coder-Next FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-Next-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit da6e2ed27304dd39abadd9c82ef50e8de67bdd4c, the API records an Apache-2.0 text-generation repo with qwen3_next and fp8 tags plus safetensors parameters BF16 683691264, F8_E4M3 78995521536, total 79679212800. The card states that this repository contains the FP8-quantized Qwen3-Coder-Next checkpoint, with fine-grained FP8 quantization, 128 block size, 80B total parameters, 3B activated parameters, 12 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)), 16 Q heads and 2 KV heads for gated attention, 32 V heads and 16 QK heads for Gated DeltaNet, and 256K default context." }, { "label": "Qwen3-Coder-Next FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8/raw/da6e2ed27304dd39abadd9c82ef50e8de67bdd4c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, FP8 quantization with dynamic activation quantization and 128x128 weight blocks, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, and 262144 max position embeddings." }, { "label": "Qwen3-Coder-Next BF16 and Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited architecture fields between Qwen/Qwen3-Coder-Next-FP8, Qwen/Qwen3-Coder-Next, and Qwen/Qwen3-Coder-Next-Base at commit 1b6df59d5f75ab51edb9ad8cb3ea69c5d0aedd57. The FP8 repo adds quantization_config while preserving the BF16 model geometry." }, { "label": "Qwen3-Coder-Next FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8/raw/da6e2ed27304dd39abadd9c82ef50e8de67bdd4c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 40 indexed shards. Stored tensors sum to 80.362904064 GB across 148383 tensors, matching the index total_size: BF16 1.367382528 GB and F8_E4M3 78.995521536 GB. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 79.740574208 GB. The resident-only input embedding is 0.622329856 GB. Routed expert tensors sum to 77.318848512 GB, or 0.151013376 GB per expert index across all 48 layers. Fixed ordinary text traffic sums to 2.421725696 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card text, served FP8 config, direct BF16 and Base config comparisons, range-read safetensors shard headers, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident input-embedding weight from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-coder-next", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Coder-Next", "title": "Qwen3-Coder-Next BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 Qwen3-Coder-Next repo.", "model_family": "qwen3-next-moe-coder", "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 159.348782592, "main_resident_weight_gb": 158.726452736, "auxiliary_resident_weight_gb": 0.622329856, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.301989888, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept as a full matrix for each ordinary generated text token", "shared_expert_notes": "The config records shared_expert_intermediate_size 512 and the model card states one shared expert. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded 80B/3B model-card parameters. Routed expert tensors are byte-uniform across all 512 expert indexes, so expected-distinct routing can use the exact per-expert stored byte group." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary text decode with speculative decoding, tool parsing, and code-agent scaffolding outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-sglang-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert compute, recurrent-state writes, and prefill scheduling are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and safetensors headers record only BF16 stored tensors. The vLLM and SGLang examples do not request a quantized KV cache, so the profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Coder-Next model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-Next", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb, the API records an Apache-2.0 text-generation repo with qwen3_next tags plus safetensors parameters BF16 79674391296. Current downloads were 1,272,766 during audit. The card states 80B total parameters, 3B activated parameters, 79B non-embedding parameters, 48 layers in a 12 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)) layout, 16 Q heads and 2 KV heads for gated attention, 32 V heads and 16 QK heads for Gated DeltaNet, 512 experts, 10 activated experts, 1 shared expert, and 262144 native context." }, { "label": "Qwen3-Coder-Next config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, bfloat16 dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, and 262144 max position embeddings." }, { "label": "Qwen3-Coder-Next safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/resolve/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 159348782592 bytes. Range-read shard headers across all 40 shards found 74391 tensors, matching the index tensor count, and stored tensors sum exactly to 159.348782592 GB, all BF16. The resident-only input embedding is 0.622329856 GB. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 158.726452736 GB. Routed expert tensors sum to 154.618822656 GB, exactly 0.301989888 GB per expert index across 512 experts. Fixed ordinary text traffic including linear attention, self attention, shared expert, router, norms, and lm_head sums to 4.10763008 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, model card text, served config, range-read safetensors shard headers, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It deliberately separates resident input-embedding weight from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-next-80b-a3b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Next-80B-A3B-Instruct-FP8", "title": "Qwen3 Next 80B A3B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-Next 80B A3B Instruct repo.", "model_family": "qwen3-next-moe-instruct", "base_model_proof": { "base_model": "Qwen/Qwen3-Next-80B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card statement and direct config comparison", "config_compatible": true, "notes": "The model card identifies this repository as the FP8-quantized Qwen3-Next-80B-A3B-Instruct checkpoint. Manual comparison found matching audited architecture fields between the FP8 repo and Qwen/Qwen3-Next-80B-A3B-Instruct; the FP8 artifact adds quantization_config while preserving the BF16 model geometry." }, "architecture": { "canonical_architecture_id": "qwen3-next-80b-a3b-instruct", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 82.03286912, "main_resident_weight_gb": 79.750217216, "auxiliary_resident_weight_gb": 2.282651904, "fixed_weight_gb": 2.42193152, "routed_expert_weight_gb": 0.151031808, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding and top-level MTP tensors", "auxiliary_scope": "model.embed_tokens.weight and top-level mtp tensors are resident for the package but not swept for ordinary non-speculative text decode", "shared_expert_notes": "The config records shared_expert_intermediate_size 512 and the model card states one shared expert. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 80B/3B model-card parameters. Routed expert tensors are byte-uniform across all 512 expert indexes, including FP8 weights and F32 scale-inverse tensors. Top-level MTP tensors are charged as resident package bytes but ordinary non-speculative decode does not sweep them." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary non-speculative text decode; MTP/speculative decoding requires a separate workload/profile path." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.008644861674296, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-sglang-fp8-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored FP8/BF16/F32 safetensors bytes. FP8 dequantization, dynamic activation quantization, router compute, expert compute, MTP speculative execution, recurrent-state writes, and prefill scheduling are outside this memory-side bound.", "notes": "The model card documents fine-grained FP8 weight quantization with block size 128. The config has no KV-cache quantization scheme and the vLLM/SGLang examples do not request FP8 KV cache, so this profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Next 80B A3B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit c5f5f263bdd5cc134092897864e8905d8fe7b928, the API records an Apache-2.0 text-generation repo with qwen3_next, fp8, endpoints_compatible, deploy:azure, and region:us tags plus safetensors parameters F32 4921664, BF16 688319744, F8_E4M3 80636542976, total 81329784384. Current downloads were 476230 when audited. The card states that this repository contains the FP8-quantized Qwen3-Next-80B-A3B-Instruct checkpoint, with fine-grained FP8 quantization, 128 block size, 80B total parameters, 3B activated parameters, 12 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)), 16 Q heads and 2 KV heads for gated attention, 32 V heads and 16 QK heads for Gated DeltaNet, 512 experts, 10 activated experts, 1 shared expert, 262144 native context, and MTP speculative-serving examples." }, { "label": "Qwen3-Next 80B A3B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8/raw/c5f5f263bdd5cc134092897864e8905d8fe7b928/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, FP8 quantization with dynamic activation quantization and 128x128 weight blocks, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, vocab size 151936, and 262144 max position embeddings." }, { "label": "Qwen3-Next 80B A3B Instruct BF16 config comparison", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/raw/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in the audited architecture fields between Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 and Qwen/Qwen3-Next-80B-A3B-Instruct at commit 9c7f2fbe84465e40164a94cc16cd30b6999b0cc7. The BF16 base API reports safetensors parameters BF16 81324862720 and the FP8 repo adds quantization_config while preserving the checked model geometry." }, { "label": "Qwen3-Next 80B A3B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8/resolve/c5f5f263bdd5cc134092897864e8905d8fe7b928/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 8 indexed shards. Stored tensors sum to 82.032869120 GB across 151479 tensors: F8_E4M3 80.636542976 GB, BF16 1.376639488 GB, and F32 0.019686656 GB. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 79.750217216 GB. Resident-only auxiliary tensors sum to 2.282651904 GB: model.embed_tokens.weight is 0.622329856 GB and top-level mtp tensors are 1.660322048 GB. Routed expert tensors sum to 77.328285696 GB, or 0.151031808 GB per expert index across all 512 experts. Fixed ordinary text traffic sums to 2.421931520 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card text, served FP8 config, direct BF16 base config comparison, range-read safetensors shard headers, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile is for ordinary non-speculative text decode bounds. It deliberately separates resident input-embedding and MTP weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-next-80b-a3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Next-80B-A3B-Instruct", "title": "Qwen3 Next 80B A3B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the official BF16 Qwen3-Next 80B A3B Instruct repo.", "model_family": "qwen3-next-moe-instruct", "architecture": { "canonical_architecture_id": "qwen3-next-80b-a3b-instruct", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 162.64972544, "main_resident_weight_gb": 158.726452736, "auxiliary_resident_weight_gb": 3.923272704, "fixed_weight_gb": 4.10763008, "routed_expert_weight_gb": 0.301989888, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding and top-level MTP tensors", "auxiliary_scope": "model.embed_tokens.weight and top-level mtp tensors are resident for the package but not swept for ordinary non-speculative text decode", "shared_expert_notes": "The config records shared_expert_intermediate_size 512 and the model card states one shared expert. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived BF16 bytes are used instead of rounded 80B/3B model-card parameters. Routed expert tensors are byte-uniform across all 512 expert indexes. Top-level MTP tensors are charged as resident package bytes but ordinary non-speculative decode does not sweep them." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config records full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state remains activation-side BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3NextForCausalLM is a text-only MoE language model. This profile models ordinary non-speculative text decode; MTP/speculative decoding requires a separate workload/profile path." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-sglang-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert compute, MTP speculative execution, recurrent-state writes, and prefill scheduling are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and safetensors headers record only BF16 stored tensors. The vLLM and SGLang examples do not request a quantized KV cache, so the profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "Qwen3-Next 80B A3B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Next-80B-A3B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit 9c7f2fbe84465e40164a94cc16cd30b6999b0cc7, the API records an Apache-2.0 text-generation repo with qwen3_next, endpoints_compatible, deploy:azure, region:us, and safetensors parameters BF16 81324862720. Current downloads were 243964 when audited. The card states 80B total parameters, 3B activated parameters, 79B non-embedding parameters, 48 layers in a 12 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)) layout, 16 Q heads and 2 KV heads for gated attention, 32 V heads and 16 QK heads for Gated DeltaNet, 512 experts, 10 activated experts, 1 shared expert, 262144 native context, and MTP speculative-serving examples for SGLang and vLLM." }, { "label": "Qwen3-Next 80B A3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/raw/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The config records Qwen3NextForCausalLM, qwen3_next, bfloat16 dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, vocab size 151936, and 262144 max position embeddings." }, { "label": "Qwen3-Next 80B A3B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/resolve/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The resolved safetensors index metadata reports total_size 162649725440 bytes. Range-read shard headers across all 41 shards found 75944 tensors, matching the index tensor count, and stored tensors sum exactly to 162.649725440 GB, all BF16. Ordinary text resident tensors, defined as model.layers plus model.norm.weight plus lm_head.weight, sum to 158.726452736 GB. Resident-only auxiliary tensors sum to 3.923272704 GB: model.embed_tokens.weight is 0.622329856 GB and top-level mtp tensors are 3.300942848 GB. Routed expert tensors sum to 154.618822656 GB, exactly 0.301989888 GB per expert index across 512 experts. Fixed ordinary text traffic including linear attention, self attention, shared expert, router, norms, and lm_head sums to 4.107630080 GB." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card text, served config, range-read safetensors shard headers, comparison with the existing FP8 Instruct and BF16 Coder-Next profiles, and the Transformers qwen3_next runtime implementation." }, "notes": "This profile is expected to be resident_not_fit on 128GB local systems. It deliberately separates resident input-embedding and MTP weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "qwen--qwen3-omni-30b-a3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Omni-30B-A3B-Instruct", "title": "Qwen3 Omni 30B A3B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3 Omni Instruct checkpoint with the full Thinker/Talker package resident.", "model_family": "qwen3-omni-moe", "architecture": { "canonical_architecture_id": "qwen3-omni-30b-a3b-instruct", "max_context_tokens": 65536, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 70.51963709, "main_resident_weight_gb": 60.44243968, "auxiliary_resident_weight_gb": 10.07719741, "fixed_weight_gb": 2.460381184, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_full_instruct_package", "traffic_scope": "ordinary text decode through the Thinker text decoder and thinker.lm_head, excluding resident-only input embedding, audio tower, vision tower, Talker, and code2wav tensors", "auxiliary_scope": "thinker.model.embed_tokens.weight, thinker.audio_tower tensors, thinker.visual tensors, all talker tensors, and all code2wav tensors are resident for the full Instruct checkpoint but not swept as full matrices for each generated text token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. The routed Thinker text experts are byte-uniform across 128 expert indexes and 48 MoE layers. The profile charges the full Instruct package as resident because the catalog row points at the Instruct checkpoint; a thinker-only or talker-disabled deployment should be represented by a separate profile/row." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Thinker text config records sliding_window null and use_sliding_window false, so text decode charges full-context K and V streams for all 48 Thinker text layers." }, "notes": "Qwen3OmniMoeForConditionalGeneration uses a MoE Thinker/Talker design. This profile models ordinary text decode after any multimodal prefill with the full Instruct checkpoint resident. Audio/video preprocessing, image/audio encoder prefill, Talker speech-output decoding, and code2wav waveform generation are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-transformers-bf16-thinker-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert compute, multimodal encoder compute, Talker speech generation, code2wav waveform synthesis, and KV writes are outside Bounds Engine v1.", "notes": "The repo config and safetensors headers record BF16 weights. The model card vLLM examples serve BF16 and tune max-model-len for memory." }, "evidence": [ { "label": "Qwen3 Omni Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Omni-30B-A3B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving", "thinker_talker_scope" ], "notes": "At commit 26291f793822fb6be9555850f06dfe95f2d7e695, the HF API reports an any-to-any Transformers repo with qwen3_omni_moe tags, safetensors BF16 parameters totaling 35,259,818,545, and used storage about 70.5 GB. The card identifies the Instruct model as containing both Thinker and Talker for audio/video/text input with audio and text output, documents disable_talker() as a memory-saving option for text-only output, lists vLLM BF16 serving examples, and records BF16 memory table entries for Instruct and Thinking variants." }, { "label": "Qwen3 Omni Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/raw/26291f793822fb6be9555850f06dfe95f2d7e695/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The served config records Qwen3OmniMoeForConditionalGeneration with Thinker text config qwen3_omni_moe_text: 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, max_position_embeddings 65536, sliding_window null, use_sliding_window false, vocab_size 152064, and BF16 dtype. It also records a Thinker audio encoder, Thinker vision encoder, a 20-layer Talker MoE text model, a 5-layer Talker code predictor, and a code2wav module." }, { "label": "Qwen3 Omni Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/resolve/26291f793822fb6be9555850f06dfe95f2d7e695/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across all 15 indexed shards. Stored tensors sum to 70.51963709 GB across 28,010 BF16 tensors, matching the index total_size. Thinker text tensors including lm_head sum to 61.065293824 GB; thinker.model.embed_tokens.weight is 0.622854144 GB and is resident-only for ordinary decode. Ordinary Thinker text swept tensors sum to 60.44243968 GB. Routed Thinker expert tensors sum to 57.982058496 GB, or exactly 0.452984832 GB per expert index. Fixed ordinary text traffic including lm_head and non-expert layer tensors sums to 2.460381184 GB. Resident-only auxiliary tensors sum to 10.07719741 GB: input embedding, audio tower, vision tower, Talker, and code2wav." }, { "label": "Qwen3 Omni model card serving notes", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct", "source_type": "manual_review", "supports": [ "serving", "thinker_talker_scope", "max_context_tokens" ], "notes": "Manual review found that the card recommends vLLM serving, shows single-GPU Instruct serving at --max-model-len 32768 and multi-GPU Instruct serving at --max-model-len 65536, notes that vLLM serve currently supports only the thinker path, and describes disabling the Talker for users who do not need audio output." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from HF API metadata, the model card, served config, safetensors index, direct range-read shard header byte grouping, and comparison with the existing Qwen3 30B A3B BF16 text MoE profile." }, "notes": "This profile is for ordinary text-output decode bounds on the full Instruct checkpoint. It does not estimate audio-output Talker throughput or code2wav waveform synthesis." }, { "id": "qwen--qwen3-omni-30b-a3b-thinking", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-Omni-30B-A3B-Thinking", "title": "Qwen3 Omni 30B A3B Thinking BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3 Omni Thinking checkpoint with the Thinker package resident.", "model_family": "qwen3-omni-moe", "architecture": { "canonical_architecture_id": "qwen3-omni-30b-a3b-thinking", "max_context_tokens": 65536, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 63.438410976, "main_resident_weight_gb": 60.44243968, "auxiliary_resident_weight_gb": 2.995971296, "fixed_weight_gb": 2.460381184, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_thinking_thinker_package", "traffic_scope": "ordinary text decode through the Thinker text decoder and thinker.lm_head, excluding resident-only input embedding, audio tower, and vision tower tensors", "auxiliary_scope": "thinker.model.embed_tokens.weight, thinker.audio_tower tensors, and thinker.visual tensors are resident for the multimodal Thinking checkpoint but not swept as full matrices for each generated text token", "notes": "Header-derived BF16 bytes are used instead of rounded model-card parameters. The routed Thinker text experts are byte-uniform across 128 expert indexes and 48 MoE layers. The Thinking checkpoint contains the Thinker component for text output and does not include Talker/code2wav tensors, unlike the Instruct checkpoint." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Thinker text config records sliding_window null and use_sliding_window false, so text decode charges full-context K and V streams for all 48 Thinker text layers." }, "notes": "Qwen3OmniMoeForConditionalGeneration uses a MoE Thinker design in this Thinking checkpoint. This profile models ordinary text decode after any multimodal prefill. Audio/video preprocessing, image/audio encoder prefill, and chain-of-thought token policy are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-or-transformers-bf16-thinker-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, router compute, expert compute, multimodal encoder compute, KV writes, and output-token policy differences for thinking traces are outside Bounds Engine v1.", "notes": "The repo config and safetensors headers record BF16 weights. The model card identifies the Thinking checkpoint as Thinker-only with text output." }, "evidence": [ { "label": "Qwen3 Omni Thinking model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Omni-30B-A3B-Thinking", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving", "thinker_scope" ], "notes": "At commit 2f443cfc4c54b14a815c0e2bb9a9d6cbcd9a748b, the HF API reports an any-to-any Transformers repo with qwen3_omni_moe tags, safetensors BF16 parameters totaling 31,719,205,488, 347831 downloads, and region:us. The card identifies the Thinking model as containing the thinker component, equipped with chain-of-thought reasoning, supporting audio/video/text input, and producing text output." }, { "label": "Qwen3 Omni Thinking config", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking/raw/2f443cfc4c54b14a815c0e2bb9a9d6cbcd9a748b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The served config records Qwen3OmniMoeForConditionalGeneration with Thinker text config qwen3_omni_moe_text: 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, max_position_embeddings 65536, sliding_window null, use_sliding_window false, vocab_size 152064, and BF16 dtype inherited from the top-level config. It also records Thinker audio and vision encoders. talker_config and token2wav_config are null." }, { "label": "Qwen3 Omni Thinking safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking/resolve/2f443cfc4c54b14a815c0e2bb9a9d6cbcd9a748b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "embedding_layout" ], "notes": "Safetensors headers were range-read across all 16 indexed shards. Stored tensors sum to 63.438410976 GB across 19,743 BF16 tensors, matching the index total_size. Thinker text body tensors plus thinker.lm_head sum to 60.442439680 GB and form ordinary swept text-decode traffic. thinker.model.embed_tokens.weight is 0.622854144 GB and is resident-only for ordinary decode. Routed Thinker expert tensors sum to 57.982058496 GB, or exactly 0.452984832 GB per expert index. Fixed ordinary text traffic including lm_head and non-expert layer tensors sums to 2.460381184 GB. Resident-only auxiliary tensors sum to 2.995971296 GB: input embedding, audio tower, and vision tower." }, { "label": "Qwen3 Omni model card serving notes", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking", "source_type": "manual_review", "supports": [ "serving", "thinker_scope", "max_context_tokens" ], "notes": "Manual review found that the shared Qwen3-Omni card describes the Thinking checkpoint as Thinker-only with text output and documents multimodal input support. Bounds Engine v1 models text-output decode only, not audio/video encoder prefill cost or the amount of thinking trace generated by a request." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from HF API metadata, the model card, served config, safetensors index, direct range-read shard header byte grouping, and comparison with the existing Qwen3 Omni Instruct profile." }, "notes": "This profile is for ordinary text-output decode bounds on the Thinking checkpoint. It does not estimate multimodal encoder prefill or make assumptions about how many hidden reasoning tokens a workload asks the model to emit." }, { "id": "qwen--qwen3-tts-12hz-1-7b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "Qwen/Qwen3-TTS-12Hz-1.7B-Base", "title": "Qwen3 TTS 12Hz 1.7B Base BF16/F32", "summary": "Unsupported profile stub for the mixed BF16/F32 Qwen3-TTS 12Hz 1.7B Base repo.", "model_family": "qwen3-tts", "architecture": { "canonical_architecture_id": "qwen3-tts-12hz-1-7b-base", "max_context_tokens": null, "weight_adapter": { "kind": "dense", "total_params_b": 2.099234881, "parameter_scope": "main_model_and_speech_tokenizer_safetensors_headers", "notes": "Range-read safetensors headers record 1928677440 BF16 parameters / 3.85735488 GB in model.safetensors plus 170557441 F32 parameters / 0.682229764 GB in speech_tokenizer/model.safetensors. The combined repo package total is 2099234881 stored parameters / 4.539584644 GB. Production bounds are disabled because this is a TTS/audio-code generation package, not an ordinary decoder-only text model." }, "kv_adapter": { "kind": "unknown", "reason": "The config records Qwen3TTSForConditionalGeneration with a talker, code predictor, speaker encoder, and separate speech tokenizer encoder/decoder. Bounds Engine v1 has no adapter for TTS acoustic code generation, codec state traffic, speaker conditioning, or speech tokenizer encode/decode work.", "notes": "Do not reuse the talker attention geometry as an ordinary text-decode KV adapter. A future TTS adapter must separately model text conditioning, audio-code generation, codec/tokenizer state, and output audio reconstruction." }, "notes": "This profile intentionally fails closed until the engine has an audited TTS/audio-generation adapter." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.1624948618601008, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-tts-audio-generation", "dequantization_notes": "No quantized weight representation is assumed, but production bounds are disabled because the runtime state and output unit are outside Bounds Engine v1.", "notes": "The main TTS model is BF16, while the bundled speech tokenizer safetensors file is F32. weight_bytes_per_param records the exact package byte average across both safetensors files." }, "evidence": [ { "label": "Qwen3-TTS 12Hz 1.7B Base API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-TTS-12Hz-1.7B-Base", "source_type": "model_card", "supports": [ "repo", "license", "total_params_b", "serving" ], "notes": "At commit fd4b254389122332181a7c3db7f27e918eec64e3, the API reports a public Apache-2.0 repo tagged qwen3_tts with BF16 safetensors count 1928677440 for the main model and siblings that include speech_tokenizer/model.safetensors." }, { "label": "Qwen3-TTS 12Hz 1.7B Base config", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/raw/fd4b254389122332181a7c3db7f27e918eec64e3/config.json", "source_type": "config", "supports": [ "architecture", "unsupported_reason" ], "notes": "The config records Qwen3TTSForConditionalGeneration, model_type qwen3_tts, tokenizer_type qwen3_tts_tokenizer_12hz, tts_model_size 1b7, a speaker encoder, a 28-layer talker with 32768 max position embeddings, and a 5-layer code predictor with 65536 max position embeddings." }, { "label": "Qwen3-TTS model card", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base", "source_type": "model_card", "supports": [ "architecture", "unsupported_reason" ], "notes": "The card describes Qwen3-TTS as text-to-speech with a discrete multi-codebook LM architecture, streaming generation, voice cloning for the Base model, and a separate Qwen3-TTS-Tokenizer-12Hz speech tokenizer." }, { "label": "Qwen3-TTS 12Hz 1.7B Base safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/resolve/fd4b254389122332181a7c3db7f27e918eec64e3/model.safetensors", "source_type": "derived_calculation", "supports": [ "total_params_b", "serving", "unsupported_reason" ], "notes": "Range-reading model.safetensors found a 58856-byte header, 480 BF16 tensors, and 1928677440 parameters / 3.85735488 GB. Range-reading speech_tokenizer/model.safetensors found a 63320-byte header, 496 F32 tensors, and 170557441 parameters / 0.682229764 GB." }, { "label": "Qwen3-TTS speech tokenizer config", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/raw/fd4b254389122332181a7c3db7f27e918eec64e3/speech_tokenizer/config.json", "source_type": "config", "supports": [ "architecture", "unsupported_reason" ], "notes": "The speech tokenizer config records Qwen3TTSTokenizerV2Model with separate encoder and decoder configs, 24 kHz audio, 12.5 Hz frame rate, vector quantization, and sliding-window audio-side state." } ], "unsupported_reason": "Qwen3-TTS is a text-to-speech/audio-code generation package with talker, speaker encoder, code predictor, and speech tokenizer components. Bounds Engine v1 only supports audited decoder-side text token memory traffic, so production tok/s bounds are disabled until a TTS adapter exists.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after the engine supports Qwen3-TTS-style audio generation and tokenizer state traffic." }, { "id": "qwen--qwen3-vl-235b-a22b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-235B-A22B-Instruct-FP8", "title": "Qwen3 VL 235B A22B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 235B A22B Instruct repo.", "model_family": "qwen3-vl-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-235B-A22B-Instruct", "relation": "quantized", "source": "Model card, served config comparison, BF16 base profile, and safetensors header review", "config_compatible": true, "notes": "The model card states that this repository contains an FP8 quantized version of Qwen/Qwen3-VL-235B-A22B-Instruct. Manual comparison found the same checked text and vision geometry as the already audited BF16 base profile; the FP8 repo adds quantization_config while preserving the served Qwen3-VL MoE geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-235b-a22b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 237.598232032, "main_resident_weight_gb": 235.200795648, "auxiliary_resident_weight_gb": 2.397436384, "fixed_weight_gb": 8.048956416, "routed_expert_weight_gb": 1.774623744, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_payloads", "traffic_scope": "ordinary text decode excludes resident visual tensors and input embeddings, and charges fixed language/logit tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config does not record a shared expert. Router/gate tensors are BF16 and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "The swept fixed subset includes non-expert model.language_model tensors except input embeddings, plus lm_head.weight. Routed expert tensors are stored as FP8 down_proj/gate_up_proj tensors plus F32 scale-inverse tensors. All 128 expert indexes are byte-uniform across the 94 layers." }, "kv_adapter": { "kind": "full_context", "layers": 94, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 94 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0081207753001782, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-qwen3-vl-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored F8_E4M3 matrix payloads plus BF16 visual/embed/lm_head/router tensors and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card recommends vLLM or SGLang for this FP8 checkpoint. The config records quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 235B A22B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 7fbcd8c9e2ad011808ed8a57d64c462605c3e381, the live API records a public non-gated Apache-2.0 image-text-to-text Transformers repo with qwen3_vl_moe, fp8, endpoints_compatible, deploy:azure, and region:us tags, 113778 downloads, and safetensors parameters F32: 14269952, BF16: 1871129328, F8_E4M3: 233798893568, total 235684292848. The model card states this is an FP8 quantized version of Qwen/Qwen3-VL-235B-A22B-Instruct with fine-grained FP8 quantization and block size 128, and recommends vLLM or SGLang because Transformers does not yet support direct loading." }, { "label": "Qwen3 VL 235B A22B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8/raw/7fbcd8c9e2ad011808ed8a57d64c462605c3e381/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "routed_experts", "routed_experts_per_token", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, root tie_word_embeddings false, 94 text layers, hidden size 4096, intermediate size 12288, MoE intermediate size 1536, 64 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, decoder_sparse_step 1, norm_topk_prob true, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and quantization_config quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and ignored visual/lm_head/router gate modules." }, { "label": "Qwen3 VL 235B A22B Instruct BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The already audited BF16 base profile records commit 710c13861be6c466e66de3f484069440b8f31389 with the same checked geometry used here: 94 text layers, 4 KV heads, 128 head dimension, 262144 context, 128 routed experts, 8 experts per token, untied embeddings, and a 27-layer visual tower." }, { "label": "Transformers 4.57.0 Qwen3 VL MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.0/src/transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "lm_head_layout", "embedding_layout" ], "notes": "Manual review for the BF16 base found Qwen3VLMoeAttention instantiates k_proj and v_proj using num_key_value_heads * head_dim and writes key_states/value_states through the cache path. Qwen3VLMoeTextSparseMoeBlock selects top_k experts from config.num_experts_per_tok. Qwen3VLMoeModel instantiates embed_tokens, and Qwen3VLMoeForConditionalGeneration instantiates a separate lm_head linear layer used for logits." }, { "label": "Qwen3 VL 235B A22B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8/resolve/7fbcd8c9e2ad011808ed8a57d64c462605c3e381/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "serving", "embedding_layout" ], "notes": "Range-read headers across all 24 safetensors shards found 1952 tensors totaling 237.598232032 GB of tensor payload: F8_E4M3 233.798893568 GB, BF16 3.742258656 GB, and F32 scale-inverse tensors 0.057079808 GB. model.language_model.embed_tokens.weight has shape [151936, 4096] and contributes 1.244659712 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.visual tensors total 1.152776672 GB. Ordinary main text/logit resident bytes total 235.200795648 GB. Fixed non-expert text/logit traffic totals 8.048956416 GB. Routed expert tensors plus scale-inverse sidecars total 227.151839232 GB, exactly 1.774623744 GB per routed expert across all 94 layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served FP8 config, audited BF16 base profile comparison, official Transformers Qwen3 VL MoE implementation, local scrape row, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token fixed language/logit weights and expected distinct routed expert traffic, and does not assume FP8 KV cache without direct serving evidence." }, { "id": "qwen--qwen3-vl-235b-a22b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-235B-A22B-Instruct", "title": "Qwen3 VL 235B A22B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 235B A22B Instruct repo.", "model_family": "qwen3-vl-moe", "architecture": { "canonical_architecture_id": "qwen3-vl-235b-a22b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 471.340045792, "main_resident_weight_gb": 468.942609408, "auxiliary_resident_weight_gb": 2.397436384, "fixed_weight_gb": 14.749817856, "routed_expert_weight_gb": 3.548381184, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode excludes resident visual tensors and input embeddings, and charges fixed language/logit tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config does not record a shared expert. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "The swept fixed subset includes non-expert model.language_model tensors except input embeddings, plus lm_head.weight. Routed expert tensors are stored as two BF16 tensors per layer with leading expert dimension 128; the profile divides their exact header bytes by 128 to obtain one routed expert across all 94 layers." }, "kv_adapter": { "kind": "full_context", "layers": 94, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 94 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The API safetensors block and range-read shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 VL 235B A22B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "max_context_tokens" ], "notes": "The card identifies Qwen3-VL-235B-A22B-Instruct as the weight repository for a Qwen3-VL MoE vision-language model, describes dense and MoE Qwen3-VL architectures, and states native 256K context expandable to 1M." }, { "label": "Qwen3 VL 235B A22B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/710c13861be6c466e66de3f484069440b8f31389/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "routed_experts", "routed_experts_per_token", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, tie_word_embeddings false, 94 text layers, hidden size 4096, 64 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, decoder_sparse_step 1, 1536 MoE intermediate size, 262144 max position embeddings, and a resident 27-layer visual tower." }, { "label": "Qwen3 VL 235B A22B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-235B-A22B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline", "revision" ], "notes": "At commit 710c13861be6c466e66de3f484069440b8f31389, the API records an Apache-2.0 image-text-to-text repo with 1676772 downloads and safetensors parameters BF16: 235670022896, total: 235670022896." }, { "label": "Qwen3 VL 235B A22B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/710c13861be6c466e66de3f484069440b8f31389/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "embedding_layout" ], "notes": "The index lists 96 safetensors shards with total_size 471340045792 bytes. Range-read shard headers record 1388 BF16 tensors totaling 235670022896 parameters / 471.340045792 GB, matching the index. model.visual tensors total 0.576388336B params / 1.152776672 GB; model.language_model.embed_tokens.weight has shape [151936, 4096] and totals 0.622329856B params / 1.244659712 GB; lm_head.weight is a separate untied tensor of the same shape and remains in swept decode traffic. Ordinary fixed text/logit traffic totals 14.749817856 GB. Routed expert tensors are model.language_model.layers.*.mlp.experts.{down_proj,gate_up_proj}: two tensors per layer for all 94 layers, 454.192791552 GB total, exactly 3.548381184 GB per routed expert across all layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card, HF API metadata, local scrape row, and direct safetensors header range reads for all 96 shards." }, "notes": "This profile is for ordinary text decode bounds. On 128GB local hardware it is expected to return resident_not_fit, but the audited profile prevents fallback estimates for this large BF16 MoE/VL checkpoint." }, { "id": "qwen--qwen3-vl-2b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-2B-Instruct", "title": "Qwen3 VL 2B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 2B Instruct repo.", "model_family": "qwen3-vl-dense", "architecture": { "canonical_architecture_id": "qwen3-vl-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 2.127532032, "swept_params_b": 1.720574976, "auxiliary_resident_params_b": 0.406957056, "resident_weight_gb": 4.255064064, "swept_weight_gb": 3.441149952, "auxiliary_resident_weight_gb": 0.813914112, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model safetensors headers including tied embed_tokens", "auxiliary_scope": "model.visual tensors are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "The config records tie_word_embeddings true and the safetensors header has no lm_head.weight. model.language_model.embed_tokens.weight is therefore the tied output projection and remains in swept ordinary text decode traffic. The auxiliary resident subset contains only model.visual tensors. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 28 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text config records dtype bfloat16, and the range-read safetensors header records only BF16 tensors." }, "evidence": [ { "label": "Qwen3 VL 2B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-2B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family", "total_params_b", "weight_format" ], "notes": "At commit 89644892e4d85e24eaac8bacfd4f463576704203, the API records an Apache-2.0 image-text-to-text repo with safetensors parameters BF16: 2127532032. The model card identifies this as the Qwen3-VL-2B-Instruct weight repository and describes native 256K context, optional 1M expansion, DeepStack visual feature fusion, and Interleaved-MRoPE." }, { "label": "Qwen3 VL 2B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/89644892e4d85e24eaac8bacfd4f463576704203/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLForConditionalGeneration, tie_word_embeddings true, BF16 text dtype, 28 text layers, hidden_size 2048, 16 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, vocab_size 151936, and a resident 24-layer visual tower with hidden_size 1024." }, { "label": "Qwen3 VL 2B Instruct safetensors header", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/resolve/89644892e4d85e24eaac8bacfd4f463576704203/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "Range-reading model.safetensors found a 76240-byte header and 625 BF16 tensors totaling 2127532032 parameters / 4.255064064 GB, matching the HF API safetensors total. model.language_model tensors sum to 1720574976 parameters / 3.441149952 GB, including model.language_model.embed_tokens.weight with shape [151936, 2048] and 311164928 parameters / 0.622329856 GB. The header has no lm_head.weight, and tie_word_embeddings is true, so the embedding table also serves as the output projection and stays in swept decode traffic. model.visual tensors sum to 406957056 parameters / 0.813914112 GB resident-only for ordinary text decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the model card, served config, HF API metadata, local scrape row, and direct range-read safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual weights from per-token swept language/logit weights while keeping the tied text embedding in swept traffic." }, { "id": "qwen--qwen3-vl-30b-a3b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", "title": "Qwen3 VL 30B A3B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 30B A3B Instruct repo.", "model_family": "qwen3-vl-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-30B-A3B-Instruct", "relation": "quantized", "source": "Model card, served config comparison, BF16 base config, and safetensors header review", "config_compatible": true, "notes": "The model card states that this repository contains an FP8 quantized version of Qwen/Qwen3-VL-30B-A3B-Instruct. Manual comparison found no differences in the checked top-level, text_config, or vision_config geometry fields between this FP8 repo and the current BF16 base repo; the FP8 repo adds quantization_config while preserving the served text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 32.251808224, "main_resident_weight_gb": 30.552215552, "auxiliary_resident_weight_gb": 1.699592672, "fixed_weight_gb": 1.554108416, "routed_expert_weight_gb": 0.226547712, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_payloads", "traffic_scope": "ordinary text decode excludes resident visual tensors and input embeddings, and charges fixed language/logit tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config does not record a shared expert. Router/gate tensors are BF16 and are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "The swept fixed subset includes non-expert model.language_model tensors except input embeddings, plus lm_head.weight. Routed expert tensors are stored with leading expert dimension 128 as FP8 down_proj/gate_up_proj tensors plus F32 scale-inverse tensors. All 128 expert indexes are byte-uniform across the 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 48 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0379508064518932, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-qwen3-vl-moe-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored F8_E4M3 matrix payloads plus BF16 visual/embed/lm_head/router tensors and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card recommends vLLM or SGLang for this FP8 checkpoint. The config records quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 30B A3B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit d9748a51ae66354c4dad665aab2c71f26cf2c8cd, the live API records an Apache-2.0 image-text-to-text Transformers repo with qwen3_vl_moe, fp8, endpoints_compatible, deploy:azure, and region:us tags, 515645 downloads, and safetensors parameters F32: 1824768, BF16: 1173755120, F8_E4M3: 29896998912, total 31072578800. The model card states this is an FP8 quantized version of Qwen/Qwen3-VL-30B-A3B-Instruct with fine-grained FP8 quantization and block size 128, and recommends vLLM or SGLang because Transformers does not yet support direct loading." }, { "label": "Qwen3 VL 30B A3B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8/raw/d9748a51ae66354c4dad665aab2c71f26cf2c8cd/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "routed_experts", "routed_experts_per_token", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, root tie_word_embeddings false, 48 text layers, hidden size 2048, intermediate size 6144, MoE intermediate size 768, 32 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, decoder_sparse_step 1, norm_topk_prob true, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and quantization_config quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and ignored visual/lm_head/router gate modules." }, { "label": "Qwen3 VL 30B A3B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "The current BF16 base API records commit 9c4b90e1e4ba969fd3b5378b57d966d725f1b86c. Manual config comparison found matching checked top-level, text, and vision geometry between the BF16 base and this FP8 artifact." }, { "label": "Transformers 4.57.0 Qwen3 VL MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.0/src/transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "lm_head_layout", "embedding_layout" ], "notes": "Manual review for the already audited BF16 base found Qwen3VLMoeAttention instantiates k_proj and v_proj using num_key_value_heads * head_dim and writes key_states/value_states through the cache path. Qwen3VLMoeTextSparseMoeBlock selects top_k experts from config.num_experts_per_tok. Qwen3VLMoeModel instantiates embed_tokens, and Qwen3VLMoeForConditionalGeneration instantiates a separate lm_head linear layer used for logits." }, { "label": "Qwen3 VL 30B A3B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8/resolve/d9748a51ae66354c4dad665aab2c71f26cf2c8cd/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "serving", "embedding_layout" ], "notes": "Range-read headers across all four safetensors shards found 1170 tensors totaling 32.251808224 GB: F8_E4M3 29.896998912 GB, BF16 2.347510240 GB, and F32 scale-inverse tensors 0.007299072 GB. model.language_model.embed_tokens.weight has shape [151936, 2048] and contributes 311164928 parameters / 0.622329856 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.visual tensors total 538631408 parameters / 1.077262816 GB. Ordinary main text/logit resident bytes total 30.552215552 GB. Fixed non-expert text/logit traffic totals 1.554108416 GB. Routed expert tensors total 28.998107136 GB, exactly 0.226547712 GB per routed expert across all 48 layers." }, { "label": "Qwen3 VL 30B A3B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct", "source_type": "manual_review", "supports": [ "architecture_compatibility", "kv_adapter" ], "notes": "The already audited BF16 base profile records the same full-context text KV geometry used here: 48 text layers, 4 KV heads, 128 head dimension, 262144 context, 128 routed experts, 8 experts per token, untied embeddings, and a 27-layer visual tower." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served FP8 config, current BF16 base config comparison, audited BF16 base profile comparison, official Transformers Qwen3 VL MoE implementation, local scrape row, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token fixed language/logit weights and expected distinct routed expert traffic, and does not assume FP8 KV cache without direct serving evidence." }, { "id": "qwen--qwen3-vl-30b-a3b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-30B-A3B-Instruct", "title": "Qwen3 VL 30B A3B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 30B A3B Instruct repo.", "model_family": "qwen3-vl-moe", "architecture": { "canonical_architecture_id": "qwen3-vl-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 62.141508064, "main_resident_weight_gb": 60.441915392, "auxiliary_resident_weight_gb": 1.699592672, "fixed_weight_gb": 2.459856896, "routed_expert_weight_gb": 0.452984832, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode excludes resident visual tensors and input embeddings, and charges fixed language/logit tensors plus expected distinct routed expert tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "shared_expert_notes": "The text config does not record a shared expert. Router/gate tensors are included in fixed_weight_gb because they are always read for MoE routing.", "notes": "The swept fixed subset includes non-expert model.language_model tensors except input embeddings, plus lm_head.weight. Routed expert tensors are stored as two BF16 tensors per layer with leading expert dimension 128; the profile divides their exact header bytes by 128 to obtain one routed expert across all 48 layers." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 48 layers, 4 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLMoeForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-qwen3-vl-moe-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, kernels, scheduler behavior, and vision encoder throughput are outside Bounds Engine v1.", "notes": "The API safetensors block and range-read shard headers record only BF16 tensors. KV traffic is charged from the BF16 text geometry." }, "evidence": [ { "label": "Qwen3 VL 30B A3B Instruct model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-30B-A3B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "model_family", "max_context_tokens", "serving", "total_params_b" ], "notes": "At repo SHA 9c4b90e1e4ba969fd3b5378b57d966d725f1b86c, the API records a public/non-gated Apache-2.0 image-text-to-text repo with qwen3_vl_moe, endpoints_compatible, deploy:azure, and region:us tags. Current downloads are 566869. The API safetensors block reports BF16: 31070754032 and total: 31070754032. The model card identifies the repo as a Qwen3-VL MoE vision-language model." }, { "label": "Qwen3 VL 30B A3B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "routed_experts", "routed_experts_per_token", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLMoeForConditionalGeneration, model_type qwen3_vl_moe, tie_word_embeddings false, BF16 text dtype, 48 text layers, hidden size 2048, 32 attention heads, 4 KV heads, 128 head dimension, 128 experts, 8 experts per token, decoder_sparse_step 1, 768 MoE intermediate size, norm_topk_prob true, 262144 max position embeddings, vocabulary size 151936, and a resident 27-layer visual tower." }, { "label": "Transformers 4.57.0 Qwen3 VL MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.0/src/transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "routed_experts", "routed_experts_per_token", "lm_head_layout", "embedding_layout" ], "notes": "Manual review found Qwen3VLMoeAttention instantiates k_proj and v_proj using num_key_value_heads * head_dim and writes key_states/value_states through the cache path. Qwen3VLMoeTextSparseMoeBlock sets top_k from config.num_experts_per_tok and selects experts with torch.topk. Qwen3VLMoeModel instantiates embed_tokens, and Qwen3VLMoeForConditionalGeneration instantiates a separate lm_head linear layer used for logits." }, { "label": "Qwen3 VL 30B A3B Instruct safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/resolve/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "embedding_layout" ], "notes": "The index records total_size 62141508064 bytes across thirteen shards. Range-read shard headers found 882 BF16 tensors totaling 31070754032 parameters / 62.141508064 GB, matching the index total. model.visual tensors total 0.538631408B params / 1.077262816 GB; model.language_model.embed_tokens.weight has shape [151936, 2048] and totals 0.311164928B params / 0.622329856 GB; lm_head.weight is a separate untied tensor of the same shape and remains in swept decode traffic. Ordinary fixed text/logit traffic totals 2.459856896 GB. Routed expert tensors are model.language_model.layers.*.mlp.experts.{down_proj,gate_up_proj}: two tensors per layer for all 48 layers, 57.982058496 GB total, exactly 0.452984832 GB per routed expert across all layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, immutable config, official Transformers 4.57.0 Qwen3 VL MoE implementation, safetensors index, and direct safetensors shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token fixed language/logit traffic and expected distinct routed expert traffic." }, { "id": "qwen--qwen3-vl-32b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-32B-Instruct-FP8", "title": "Qwen3 VL 32B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 32B Instruct repo.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-32B-Instruct", "relation": "quantized", "source": "Model card, served config comparison, BF16 Instruct config, and safetensors header review", "config_compatible": true, "notes": "The model card states that this repository contains an FP8 quantized version of Qwen/Qwen3-VL-32B-Instruct. Manual comparison found no differences in the checked top-level, text_config, or vision_config geometry fields between this FP8 repo and the current BF16 Instruct repo; the FP8 repo adds quantization_config while preserving the served text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-32b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.359294704, "swept_params_b": 31.986115584, "auxiliary_resident_params_b": 1.37317912, "resident_weight_gb": 35.516776928, "swept_weight_gb": 32.770418688, "auxiliary_resident_weight_gb": 2.74635824, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_payloads", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "The swept subset includes model.language_model layer tensors, model.language_model.norm.weight, F32 FP8 scale-inverse tensors, and lm_head.weight. The root config records tie_word_embeddings false and the safetensors headers store a separate BF16 lm_head.weight, so model.language_model.embed_tokens.weight is resident-only for ordinary decode while lm_head.weight remains swept output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 64 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0646740958747338, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored F8_E4M3 matrix payloads plus BF16 visual/embed/lm_head tensors and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card recommends vLLM or SGLang for this FP8 checkpoint. The config records quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 32B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-32B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 4bf2c2f39c37c0fede78bede4056e1f18cdf8109, the API records an Apache-2.0 image-text-to-text Transformers repo with qwen3_vl, fp8, endpoints_compatible, deploy:azure, and region:us tags, 508282 downloads, and safetensors parameters F32: 1904640, BF16: 2151768304, F8_E4M3: 31205621760, total 33359294704. The model card states this is an FP8 quantized version of Qwen/Qwen3-VL-32B-Instruct with fine-grained FP8 quantization and block size 128, and recommends vLLM or SGLang because Transformers does not yet support direct loading." }, { "label": "Qwen3 VL 32B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8/raw/4bf2c2f39c37c0fede78bede4056e1f18cdf8109/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLForConditionalGeneration, root tie_word_embeddings false, 64 text layers, hidden size 5120, intermediate size 25600, 64 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and quantization_config quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and 245 ignored visual/lm_head modules." }, { "label": "Qwen3 VL 32B Instruct BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct/raw/0cfaf48183f594c314753d30a4c4974bc75f3ccb/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "The current BF16 Instruct base API records commit 0cfaf48183f594c314753d30a4c4974bc75f3ccb and safetensors BF16 parameters 33357390064. Manual config comparison found matching checked top-level, text, and vision geometry between the BF16 Instruct base and this FP8 artifact." }, { "label": "Qwen3 VL 32B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct", "source_type": "manual_review", "supports": [ "architecture_compatibility", "kv_adapter" ], "notes": "The already audited Qwen/Qwen3-VL-32B-Instruct profile records the same full-context text KV geometry used here: 64 text layers, hidden size 5120, 8 KV heads, 128 head dimension, 262144 context, untied embeddings, and a 27-layer visual tower." }, { "label": "Qwen3 VL 32B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8/resolve/4bf2c2f39c37c0fede78bede4056e1f18cdf8109/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "Range-read headers across all seven safetensors shards found 1506 tensors totaling 35.516776928 GB: F8_E4M3 31.205621760 GB, BF16 4.303536608 GB, and F32 scale-inverse tensors 0.007618560 GB. model.language_model.embed_tokens.weight has shape [151936, 5120] and contributes 777912320 parameters / 1.555824640 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.visual tensors total 595266800 parameters / 1.190533600 GB. Ordinary text swept traffic is 32.770418688 GB, and auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, total 1.373179120 parameters / 2.746358240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served FP8 config, current BF16 Instruct config comparison, audited BF16 Instruct profile comparison, local scrape row, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token swept language/logit weights and does not assume FP8 KV cache without direct serving evidence." }, { "id": "qwen--qwen3-vl-32b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-32B-Instruct", "title": "Qwen3 VL 32B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 32B Instruct repo.", "model_family": "qwen3-vl-dense", "architecture": { "canonical_architecture_id": "qwen3-vl-32b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.357390064, "swept_params_b": 31.984210944, "auxiliary_resident_params_b": 1.37317912, "resident_weight_gb": 66.714780128, "swept_weight_gb": 63.968421888, "auxiliary_resident_weight_gb": 2.74635824, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors and model.language_model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 64 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text config records dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 VL 32B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family" ], "notes": "The card identifies Qwen3-VL-32B-Instruct as an Apache-2.0 image-text-to-text vision-language repo and describes native 256K context plus DeepStack visual feature fusion." }, { "label": "Qwen3 VL 32B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct/raw/0cfaf48183f594c314753d30a4c4974bc75f3ccb/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLForConditionalGeneration, tie_word_embeddings false, BF16 text dtype, 64 text layers, hidden size 5120, 64 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, and a resident 27-layer visual tower." }, { "label": "Qwen3 VL 32B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-32B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit 0cfaf48183f594c314753d30a4c4974bc75f3ccb, the API safetensors block records BF16: 33357390064 and total: 33357390064, which this profile stores as 33.357390064B resident parameters." }, { "label": "Qwen3 VL 32B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct/raw/0cfaf48183f594c314753d30a4c4974bc75f3ccb/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists fourteen safetensors shards. Range-read shard headers record 1058 BF16 tensors totaling 33357390064 parameters and 66.714780128 GB, matching index total_size. model.language_model.embed_tokens.weight has shape [151936, 5120] and contributes 777912320 parameters / 1.55582464 GB. lm_head.weight is a separate untied tensor of the same shape and remains in swept decode traffic. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 31984210944 parameters / 63.968421888 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1373179120 parameters / 2.74635824 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token swept language/logit weights." }, { "id": "qwen--qwen3-vl-32b-thinking-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-32B-Thinking-FP8", "title": "Qwen3 VL 32B Thinking FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 32B Thinking repo.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-32B-Thinking", "relation": "quantized", "source": "Model card, served config comparison, BF16 Thinking config, and safetensors header review", "config_compatible": true, "notes": "The model card states that this repository contains an FP8 quantized version of Qwen/Qwen3-VL-32B-Thinking. Manual comparison found no differences in the checked top-level, text_config, or vision_config geometry fields between this FP8 repo, the BF16 Thinking repo, and the already audited BF16 Instruct sibling; the FP8 repo adds quantization_config while preserving the served text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-32b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 33.359294704, "swept_params_b": 31.986115584, "auxiliary_resident_params_b": 1.37317912, "resident_weight_gb": 35.516776928, "swept_weight_gb": 32.770418688, "auxiliary_resident_weight_gb": 2.74635824, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_payloads", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "The swept subset includes model.language_model layer tensors, model.language_model.norm.weight, F32 FP8 scale-inverse tensors, and lm_head.weight. The root config records tie_word_embeddings false and the safetensors headers store a separate BF16 lm_head.weight, so model.language_model.embed_tokens.weight is resident-only for ordinary decode while lm_head.weight remains swept output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 64 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window, recurrent-state text cache, or quantized KV cache scheme." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0646740958747338, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored F8_E4M3 matrix payloads plus BF16 visual/embed/lm_head tensors and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The model card recommends vLLM or SGLang for this FP8 checkpoint. The config records quant_method fp8, fmt e4m3, dynamic activation quantization, 128x128 weight blocks, and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 32B Thinking FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-32B-Thinking-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 3eee143c9b355ba456390568e5970e2e60f5582b, the API records an Apache-2.0 image-text-to-text Transformers repo with qwen3_vl, fp8, endpoints_compatible, deploy:azure, and region:us tags, 306899 downloads, and safetensors parameters F32: 1904640, BF16: 2151768304, F8_E4M3: 31205621760, total 33359294704. The model card states this is an FP8 quantized version of Qwen/Qwen3-VL-32B-Thinking with fine-grained FP8 quantization and block size 128, and recommends vLLM or SGLang because Transformers does not yet support direct loading." }, { "label": "Qwen3 VL 32B Thinking FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-FP8/raw/3eee143c9b355ba456390568e5970e2e60f5582b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLForConditionalGeneration, root tie_word_embeddings false, 64 text layers, hidden size 5120, intermediate size 25600, 64 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and quantization_config quant_method fp8, fmt e4m3, dynamic activation scheme, weight_block_size [128, 128], and 245 ignored visual/lm_head modules." }, { "label": "Qwen3 VL 32B Thinking BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking/raw/7edd10ffd1196091948fb245ff63e406ccb2d4d1/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "The BF16 Thinking base API records commit 7edd10ffd1196091948fb245ff63e406ccb2d4d1 and safetensors BF16 parameters 33357390064. Manual config comparison found matching checked text and vision geometry between the BF16 Thinking base and this FP8 artifact." }, { "label": "Qwen3 VL 32B Instruct BF16 profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct", "source_type": "manual_review", "supports": [ "architecture_compatibility", "kv_adapter" ], "notes": "The already audited Qwen/Qwen3-VL-32B-Instruct profile has the same checked text and vision geometry as the FP8 Thinking repo: 64 text layers, hidden size 5120, 8 KV heads, 128 head dimension, 262144 context, untied embeddings, and a 27-layer visual tower. This supports reusing the same full-context text KV geometry." }, { "label": "Qwen3 VL 32B Thinking FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-FP8/resolve/3eee143c9b355ba456390568e5970e2e60f5582b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "Range-read headers across all seven safetensors shards found 1506 tensors totaling 35.516776928 GB: F8_E4M3 31.205621760 GB, BF16 4.303536608 GB, and F32 scale-inverse tensors 0.007618560 GB. model.language_model.embed_tokens.weight has shape [151936, 5120] and contributes 777912320 parameters / 1.555824640 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.visual tensors total 595266800 parameters / 1.190533600 GB. Ordinary text swept traffic is 32.770418688 GB, and auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, total 1.373179120 parameters / 2.746358240 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served FP8 config, BF16 Thinking config comparison, audited BF16 Instruct profile comparison, local scrape row, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token swept language/logit weights and does not assume FP8 KV cache without direct serving evidence." }, { "id": "qwen--qwen3-vl-4b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-4B-Instruct-FP8", "title": "Qwen3 VL 4B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 4B Instruct repo.", "model_family": "qwen3-vl-dense-fp8", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-4B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-VL-4B-Instruct as its base model. Manual comparison found matching checked text and vision architecture fields. The FP8 repo changes root tie_word_embeddings to false, stores a separate lm_head.weight, and adds FP8 quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.826993728, "swept_params_b": 4.022689856, "auxiliary_resident_params_b": 0.804303872, "resident_weight_gb": 6.021115136, "swept_weight_gb": 4.412507392, "auxiliary_resident_weight_gb": 1.608607744, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.language_model.embed_tokens.weight input lookup and includes language layer tensors plus lm_head.weight output projection", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "The text_config still records tie_word_embeddings true, but the root config records tie_word_embeddings false and the FP8 safetensors headers store a separate BF16 lm_head.weight. Therefore model.language_model.embed_tokens.weight is resident-only for ordinary decode, while lm_head.weight remains in swept output-projection traffic. Visual tensors remain resident-only for this text-decode profile." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.2473840811255348, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-fp8-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads, BF16 visual/input/head/norm/bias tensors, and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, vision prefill, and compute overhead are outside Bounds Engine v1.", "notes": "The model card documents fine-grained FP8 quantization with block size 128. The config records quant_method fp8, fmt e4m3, dynamic activation quantization, 128x128 weight blocks, and no quantized KV cache scheme, so KV cache bytes are charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 4B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-4B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit fefbb44cbcce8d1bb7e20b920b94f77432b3446d, the API records an Apache-2.0 image-text-to-text Transformers repo with base_model Qwen/Qwen3-VL-4B-Instruct, fp8 and region:us tags, 474505 downloads, and safetensors parameters F32: 221760, BF16: 1193456128, F8_E4M3: 3633315840, total 4826993728. The model card identifies this as an FP8 quantized version of Qwen3-VL-4B-Instruct with fine-grained FP8 quantization and block size 128." }, { "label": "Qwen3 VL 4B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-FP8/raw/fefbb44cbcce8d1bb7e20b920b94f77432b3446d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof", "auxiliary_resident_scope" ], "notes": "The config records Qwen3VLForConditionalGeneration, root tie_word_embeddings false, text_config tie_word_embeddings true, 36 text layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, and a resident 24-layer visual tower. The quantization config records quant_method fp8, fmt e4m3, activation_scheme dynamic, weight_block_size [128, 128], modules_to_not_convert lm_head and model.visual, and ignored visual layers." }, { "label": "Qwen3 VL 4B Instruct BF16 base profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited BF16 Qwen3-VL-4B-Instruct profile found matching checked text and vision geometry and the same native 262144 context. Unlike the BF16 base, this FP8 package stores an untied lm_head.weight tensor and scale-inverse tensors." }, { "label": "Qwen3 VL 4B Instruct FP8 safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-FP8/resolve/fefbb44cbcce8d1bb7e20b920b94f77432b3446d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "Range-read headers across the two safetensors shards found 966 tensors totaling 6.021115136 GB: F8_E4M3 3.633315840 GB, BF16 2.386912256 GB, and F32 scale-inverse tensors 0.000887040 GB. model.visual tensors contribute 0.830695424 GB resident-only. model.language_model.embed_tokens.weight and lm_head.weight are both BF16 [151936, 2560] tensors contributing 0.777912320 GB each; the input embedding is resident-only for ordinary decode and lm_head.weight is swept. Ordinary text swept traffic is 4.412507392 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served FP8 config, base profile comparison, direct safetensors header byte grouping, and local scrape row." }, "notes": "This is a self-contained dense FP8 profile for ordinary text decode bounds. It intentionally does not assume FP8 KV cache, vision throughput, or dequantization speedups without direct serving evidence." }, { "id": "qwen--qwen3-vl-4b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-4B-Instruct", "title": "Qwen3 VL 4B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 4B Instruct repo.", "model_family": "qwen3-vl-dense", "architecture": { "canonical_architecture_id": "qwen3-vl-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.437815808, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0.415347712, "resident_weight_gb": 8.875631616, "swept_weight_gb": 8.044936192, "auxiliary_resident_weight_gb": 0.830695424, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model safetensors headers including tied embed_tokens", "auxiliary_scope": "model.visual tensors are resident for the multimodal package but not swept for each ordinary text decode token", "notes": "The config records tie_word_embeddings true and the safetensors headers have no lm_head.weight. model.language_model.embed_tokens.weight is therefore the tied output projection and remains in swept ordinary text decode traffic. The auxiliary resident subset contains only model.visual tensors. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text config records dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 VL 4B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family" ], "notes": "The card identifies Qwen3-VL-4B-Instruct as an Apache-2.0 image-text-to-text vision-language repo and describes native 256K context plus DeepStack visual feature fusion." }, { "label": "Qwen3 VL 4B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLForConditionalGeneration, tie_word_embeddings true, BF16 text dtype, 36 text layers, hidden size 2560, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, and a resident 24-layer visual tower." }, { "label": "Qwen3 VL 4B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-4B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit ebb281ec70b05090aa6165b016eac8ec08e71b17, the API safetensors block records BF16: 4437815808 and total: 4437815808, which this profile stores as 4.437815808B resident parameters." }, { "label": "Qwen3 VL 4B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/ebb281ec70b05090aa6165b016eac8ec08e71b17/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists two safetensors shards. Range-read shard headers record 713 BF16 tensors totaling 4437815808 parameters and 8.875631616 GB, matching index total_size. model.language_model.embed_tokens.weight has shape [151936, 2560] and contributes 388956160 parameters / 0.77791232 GB. The headers have no lm_head.weight. Ordinary text swept tensors, defined as all model.language_model tensors including the tied embedding/output projection, sum to 4022468096 parameters / 8.044936192 GB. Auxiliary resident tensors, defined as model.visual tensors, sum to 415347712 parameters / 0.830695424 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual weights from per-token swept language/logit weights while keeping the tied text embedding in swept traffic." }, { "id": "qwen--qwen3-vl-8b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-8B-Instruct-FP8", "title": "Qwen3 VL 8B Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the official FP8 Qwen3-VL 8B Instruct repo.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-8B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, audited BF16 base profile, and safetensors header review", "config_compatible": true, "notes": "The FP8 artifact records Qwen/Qwen3-VL-8B-Instruct as its quantized base model. Manual comparison found no differences in the architecture fields used by this profile between the FP8 config and the audited BF16 base config; the FP8 repo adds quantization_config while preserving the served text and vision geometry." }, "architecture": { "canonical_architecture_id": "qwen3-vl-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.767547632, "swept_params_b": 7.56882944, "auxiliary_resident_params_b": 1.198718192, "resident_weight_gb": 10.590175712, "swept_weight_gb": 8.192739328, "auxiliary_resident_weight_gb": 2.397436384, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_tensor_payloads", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "The swept subset includes model.language_model layer tensors, model.language_model.norm.weight, F32 FP8 scale-inverse tensors, and lm_head.weight. lm_head and model.visual are explicitly excluded from FP8 conversion by the quantization config and remain BF16. The auxiliary resident subset includes the BF16 visual tower and BF16 input embedding." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.20788345344686, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-vllm-fp8-bf16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored F8_E4M3 matrix payloads plus BF16 visual/embed/lm_head tensors and F32 scale-inverse tensors from safetensors headers. FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and quantization_config quant_method fp8, fmt e4m3, dynamic activation quantization, 128x128 weight blocks, modules_to_not_convert lm_head and model.visual, and no quantized KV cache scheme. KV cache bytes are therefore charged as BF16." }, "evidence": [ { "label": "Qwen3 VL 8B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-8B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving", "total_params_b" ], "notes": "At commit 9cdc6310a8cb770ce18efaf4e9935334512aee45, the API records a public Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen3-VL-8B-Instruct, base_model_relation quantized, qwen3_vl/fp8 tags, current downloads 710972, and safetensors parameters F32 423936, BF16 1821356272, F8_E4M3 6945767424, total 8767547632." }, { "label": "Qwen3 VL 8B Instruct FP8 config", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-FP8/raw/9cdc6310a8cb770ce18efaf4e9935334512aee45/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "base_model_proof" ], "notes": "The config records Qwen3VLForConditionalGeneration, tie_word_embeddings false, 36 text layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, 151936 vocab size, a 27-layer visual tower, and quantization_config quant_method fp8, fmt e4m3, dynamic activation scheme, 128x128 weight blocks, and modules_to_not_convert lm_head and model.visual." }, { "label": "Qwen3 VL 8B Instruct BF16 base profile and config comparison", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited BF16 Qwen3-VL-8B-Instruct profile and pinned base config found matching served architecture geometry: same top-level architecture, tie_word_embeddings false, 36 text layers, 4096 hidden size, 12288 intermediate size, 32 attention heads, 8 KV heads, 128 head dimension, 262144 context, 151936 vocab size, 27 visual layers, 1152 visual hidden size, and 4096 visual output hidden size." }, { "label": "Qwen3 VL 8B Instruct FP8 safetensors index and shard headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-FP8/raw/9cdc6310a8cb770ce18efaf4e9935334512aee45/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "serving", "embedding_layout" ], "notes": "The index lists two safetensors shards. Range-read shard headers record 1002 tensors totaling 8767547632 parameters / 10.590175712 GB: F32 423936 parameters / 0.001695744 GB, BF16 1821356272 parameters / 3.642712544 GB, and F8_E4M3 6945767424 parameters / 6.945767424 GB. model.language_model.embed_tokens.weight has shape [151936, 4096] and contributes 622329856 parameters / 1.244659712 GB resident-only for ordinary decode. lm_head.weight is separate BF16 with the same shape and remains in swept decode traffic. model.visual tensors sum to 576388336 parameters / 1.152776672 GB. Ordinary text swept tensors, defined as model.language_model layers plus model.language_model.norm.weight plus lm_head.weight, sum to 7568829440 parameters / 8.192739328 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1198718192 parameters / 2.397436384 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served FP8 config, audited BF16 base comparison, local scrape row, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token swept language/logit weights and does not assume FP8 KV cache without direct serving evidence." }, { "id": "qwen--qwen3-vl-8b-instruct", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3-VL-8B-Instruct", "title": "Qwen3 VL 8B Instruct BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Qwen3-VL 8B Instruct repo.", "model_family": "qwen3-vl-dense", "architecture": { "canonical_architecture_id": "qwen3-vl-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.767123696, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 1.198718192, "resident_weight_gb": 17.534247392, "swept_weight_gb": 15.136811008, "auxiliary_resident_weight_gb": 2.397436384, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package and input lookup but not swept for each ordinary text decode token", "notes": "The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. The auxiliary resident subset includes model.visual tensors and model.language_model.embed_tokens.weight. All stored tensors are BF16." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records full-context attention geometry with 36 layers, 8 KV heads, and a 128 head dimension. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "Qwen3VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The text config records dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3 VL 8B Instruct model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "model_family" ], "notes": "The card identifies Qwen3-VL-8B-Instruct as an Apache-2.0 image-text-to-text vision-language repo and describes native 256K context plus DeepStack visual feature fusion." }, { "label": "Qwen3 VL 8B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct/raw/0c351dd01ed87e9c1b53cbc748cba10e6187ff3b/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3VLForConditionalGeneration, tie_word_embeddings false, BF16 text dtype, 36 text layers, hidden size 4096, 32 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, and a resident 27-layer visual tower." }, { "label": "Qwen3 VL 8B Instruct Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-8B-Instruct", "source_type": "derived_calculation", "supports": [ "total_params_b", "weight_format", "license", "pipeline" ], "notes": "At commit 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b, the API safetensors block records BF16: 8767123696 and total: 8767123696, which this profile stores as 8.767123696B resident parameters." }, { "label": "Qwen3 VL 8B Instruct safetensors headers", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct/raw/0c351dd01ed87e9c1b53cbc748cba10e6187ff3b/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "embedding_layout" ], "notes": "The index lists four safetensors shards. Range-read shard headers record 750 BF16 tensors totaling 8767123696 parameters and 17.534247392 GB, matching index total_size. model.language_model.embed_tokens.weight has shape [151936, 4096] and contributes 622329856 parameters / 1.244659712 GB. lm_head.weight is a separate untied tensor of the same shape and remains in swept decode traffic. Ordinary text swept tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 7568405504 parameters / 15.136811008 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1198718192 parameters / 2.397436384 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card, HF API metadata, local scrape row, and direct safetensors header grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual and input-embedding weights from per-token swept language/logit weights." }, { "id": "qwen--qwen3guard-gen-0-6b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Qwen/Qwen3Guard-Gen-0.6B", "title": "Qwen3Guard Gen 0.6B BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 Qwen3Guard Gen 0.6B repo.", "model_family": "qwen3-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B", "relation": "finetune", "source": "Hugging Face model-card metadata, served config comparison, audited Qwen3 0.6B profile comparison, and direct safetensors header grouping", "config_compatible": false, "notes": "The model card identifies Qwen/Qwen3-0.6B as the base model. Manual comparison with the audited Qwen/Qwen3-0.6B profile found matching tensor geometry and embedding layout but a different context length: this Qwen3Guard config records 32768 max position embeddings while Qwen/Qwen3-0.6B records 40960." }, "architecture": { "canonical_architecture_id": "qwen3guard-gen-0-6b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.751632384, "swept_params_b": 0.59604992, "auxiliary_resident_params_b": 0.155582464, "resident_weight_gb": 1.503264768, "swept_weight_gb": 1.19209984, "auxiliary_resident_weight_gb": 0.311164928, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Range-read safetensors headers record 311 BF16 tensors totaling 751632384 stored parameters. The config marks tie_word_embeddings true, but the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors; this profile charges stored resident bytes and excludes only the input embedding lookup from ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config has no sliding-window setting, so the v1 profile charges full-context K and V streams for all language layers." }, "notes": "Dense Qwen3ForCausalLM safety-classification fine-tune using the served repo config and direct safetensors header grouping rather than deriving structure from the model name." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and direct safetensors headers record only BF16 tensors." }, "evidence": [ { "label": "Qwen3Guard Gen 0.6B model card", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/raw/fada3b2f655b89601929198343c94cd2f64d93cc/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The pinned model card records a public Apache-2.0 text-generation repo built upon Qwen/Qwen3-0.6B. It describes Qwen3Guard-Gen as a generative safety moderation model for prompt and response classification." }, { "label": "Qwen3Guard Gen 0.6B HF API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3Guard-Gen-0.6B", "source_type": "derived_calculation", "supports": [ "downloads", "total_params_b", "weight_format", "revision" ], "notes": "At commit fada3b2f655b89601929198343c94cd2f64d93cc, the API records 185541 downloads, a public non-gated Apache-2.0 text-generation repo with region:us, and safetensors parameters BF16 751632384, total 751632384." }, { "label": "Qwen3Guard Gen 0.6B config", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/raw/fada3b2f655b89601929198343c94cd2f64d93cc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Qwen3ForCausalLM, bfloat16, tie_word_embeddings true, hidden size 1024, intermediate size 3072, 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, rope_theta 1000000, vocab size 151936, use_sliding_window false, and no sliding_window setting." }, { "label": "Qwen3Guard Gen 0.6B safetensors header", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/resolve/fada3b2f655b89601929198343c94cd2f64d93cc/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "embedding_layout" ], "notes": "A range-read of model.safetensors found a 35552-byte header with 311 BF16 tensors totaling 751632384 parameters / 1.503264768 GB. model.embed_tokens.weight has shape [151936, 1024] and contributes 0.311164928 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 1.192099840 GB." }, { "label": "Qwen3 0.6B audited profile comparison", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "manual_review", "supports": [ "base_model_proof", "architecture_compatibility" ], "notes": "Manual comparison with the audited Qwen/Qwen3-0.6B profile found matching layers, attention heads, KV heads, head dimension, hidden size, intermediate size, rope_theta, vocab size, separate stored input embedding, and separate lm_head. Qwen3Guard records 32768 max position embeddings while the Qwen/Qwen3-0.6B profile records 40960." } ], "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned served config, pinned model card, direct range-read safetensors header, existing Qwen3 0.6B profile comparison, and local scrape row." }, "notes": "This is a self-contained dense BF16 profile for production profile-backed bounds. It uses Qwen3Guard's served 32768-token context instead of inheriting the Qwen/Qwen3-0.6B context setting." }, { "id": "readyart--dark-scarlett-v0-3-26b-a4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF", "title": "ReadyArt Dark Scarlett v0.3 26B A4B GGUF IQ4_XS", "summary": "Audited memory-side text-decode bounds profile for the API-selected i1-IQ4_XS GGUF artifact of ReadyArt Dark Scarlett v0.3 26B A4B.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "finetune", "source": "Hugging Face model card/API metadata, Google base config, selected linked-object sizes, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a finetuned GGUF distribution derived from google/gemma-4-26B-A4B-it. The selected GGUF header records the same Gemma 4 26B A4B text architecture as the Google base config." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.917726432, "main_resident_weight_gb": 13.901903808, "auxiliary_resident_weight_gb": 0.015822624, "fixed_weight_gb": 1.531313088, "routed_expert_weight_gb": 0.09664524, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "shared_expert_notes": "The GGUF header records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf. A GGUF v3 range-read found 658 tensors and 50 metadata entries. Tensor spans total 13.901903808 GB, while the linked file is 13.917726432 GB. Routed expert weight tensors total 12.370590720 GB across 30 layers and 128 expert indexes, or 0.096645240 GB per expert index. Non-expert tensor spans total 1.531313088 GB, including token_embd.weight because the selected file has no separate output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. Gemma 4 full-attention layers use K=V behavior, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile targets the API-selected main text GGUF artifact. Other quantized GGUF siblings and the imatrix file are separate artifacts and should get separate profiles if selected by a workload." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5515653344568819, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-xs-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, image handling, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the i1-IQ4_XS GGUF because HF API gguf.totalFileSize exactly matches Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "ReadyArt Dark Scarlett v0.3 26B A4B GGUF API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit b076da94ae47dd6f270dad74caa5094d6e505266, the live API records a public non-gated GGUF repo with base_model google/gemma-4-26B-A4B-it, base_model_relation finetune, Apache-2.0 license metadata, region:us, 144312 downloads, GGUF architecture gemma4, context_length 262144, gguf.total 25233142046, and gguf.totalFileSize 13917726432." }, { "label": "ReadyArt Dark Scarlett v0.3 26B A4B GGUF model card", "url": "https://huggingface.co/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF/raw/b076da94ae47dd6f270dad74caa5094d6e505266/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "runtime_format" ], "notes": "The card metadata records base_model google/gemma-4-26B-A4B-it, base_model_relation finetune, Apache-2.0 licensing, Gemma 4 tags, imatrix metadata, and mature-audience roleplay tags." }, { "label": "Google Gemma 4 26B A4B IT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The pinned Google config records Gemma4ForConditionalGeneration, bfloat16 source dtype, 30 text layers, 16 attention heads, 8 local KV heads, 5 full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, and 262144 max position embeddings." }, { "label": "ReadyArt Dark Scarlett selected GGUF linked-object HEAD checks", "url": "https://huggingface.co/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF/tree/b076da94ae47dd6f270dad74caa5094d6e505266", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF CLI info and exact linked-object HEAD checks found Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf is 13.917726432 GB, exactly matching API gguf.totalFileSize. Other main siblings range from 14.488056544 GB to 26.859859392 GB, and the imatrix file is a separate 0.056941536 GB artifact." }, { "label": "ReadyArt Dark Scarlett i1-IQ4_XS GGUF range-read tensor index", "url": "https://huggingface.co/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF/resolve/b076da94ae47dd6f270dad74caa5094d6e505266/Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 658 tensors and 50 metadata entries. Tensor spans sum to 13.901903808 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015822624 GB. Tensor spans split into IQ4_XS 8.867809280 GB, IQ4_NL 4.382484480 GB, Q6_K 0.605552640 GB, and F32 0.046057408 GB. Non-expert tensor spans total 1.531313088 GB. Routed expert weight tensors total 12.370590720 GB across 30 layers and 128 expert indexes, or 0.096645240 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, K=V full-attention geometry, separate sliding-layer K/V projections, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI model info, model card metadata, Google base config, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected i1-IQ4_XS artifact." }, "notes": "Use this profile for the API-selected ReadyArt Dark Scarlett i1-IQ4_XS GGUF text artifact. Do not infer Q4_K_M, Q5_K_M, Q6_K, Q8_0, or imatrix sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "readyart--melody1437-26b-a4b-v2-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ReadyArt/Melody1437-26B-A4B-v2.0-GGUF", "title": "ReadyArt Melody1437 26B A4B v2.0 GGUF HB16 Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the API-selected HB16 Q4_K_M GGUF artifact of ReadyArt Melody1437 26B A4B v2.0.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "ReadyArt/Melody1437-26B-A4B-v2.0", "relation": "quantized", "source": "Hugging Face model card/API metadata, gated base access check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of ReadyArt/Melody1437-26B-A4B-v2.0, which itself records google/gemma-4-26B-A4B-it as its base model. The ReadyArt base config is gated in this audit environment, so this profile uses the public selected GGUF header as the architecture source instead of claiming a direct base-config comparison." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.666858464, "main_resident_weight_gb": 17.651035256, "auxiliary_resident_weight_gb": 0.015823208, "fixed_weight_gb": 2.520870008, "routed_expert_weight_gb": 0.118204416, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "shared_expert_notes": "The GGUF header records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf. A GGUF v3 range-read found 658 tensors and 47 metadata entries. Tensor spans total 17.651035256 GB, while the linked file is 17.666858464 GB. Routed expert weight tensors total 15.130165248 GB across 30 layers and 128 expert indexes, or 0.118204416 GB per expert index. Non-expert tensor spans total 2.520870008 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. Gemma 4 full-attention layers use K=V behavior, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format." }, "notes": "This profile targets the API-selected main text GGUF artifact. Other GGUF siblings and the imatrix file are separate artifacts and should get separate profiles if selected by a workload." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.7001450089645328, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-hb16-q4-k-m-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, image handling, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the HB16 Q4_K_M GGUF because HF API gguf.totalFileSize exactly matches Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "ReadyArt Melody1437 26B A4B v2.0 GGUF API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit b22520a5c83db6ff0d12aef1fd461c932ba86fbf, the live API records a public non-gated GGUF repo with base_model ReadyArt/Melody1437-26B-A4B-v2.0, Apache-2.0 license metadata, region:us, 156079 downloads, GGUF architecture gemma4, context_length 262144, gguf.total 25233142046, and gguf.totalFileSize 17666858464." }, { "label": "ReadyArt Melody1437 26B A4B v2.0 GGUF model card", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF/raw/b22520a5c83db6ff0d12aef1fd461c932ba86fbf/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "runtime_format" ], "notes": "The card records base_model ReadyArt/Melody1437-26B-A4B-v2.0, base_model_relation quantized, Apache-2.0 licensing, and Gemma 4 roleplay/instruct tags." }, { "label": "ReadyArt Melody1437 26B A4B v2.0 base API metadata and gated config check", "url": "https://huggingface.co/api/models/ReadyArt/Melody1437-26B-A4B-v2.0", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base API metadata records a public but manually gated safetensors repo, base_model google/gemma-4-26B-A4B-it, Apache-2.0 license metadata, and safetensors total 26544131376 parameters. Raw base config access at commit 8936d5c4fa8ddf371cb10a898d6a3d932c23c941 returned HTTP 401 in this audit environment, so this profile does not claim direct base-config compatibility." }, { "label": "ReadyArt Melody1437 selected GGUF linked-object HEAD checks", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF/tree/b22520a5c83db6ff0d12aef1fd461c932ba86fbf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HF CLI dry-run and exact HEAD checks found Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf is 17.666858464 GB, exactly matching API gguf.totalFileSize. Nearby alternatives include i1-Q4_K_M_hb16 at 17.666859264 GB and non-HB16 Q4_K_M at 16.796016096 GB, so the API-selected artifact is the HB16 Q4_K_M file." }, { "label": "ReadyArt Melody1437 HB16 Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF/resolve/b22520a5c83db6ff0d12aef1fd461c932ba86fbf/Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 658 tensors and 47 metadata entries. Tensor spans sum to 17.651035256 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015823208 GB. Tensor spans split into Q4_K 9.347272704 GB, Q8_0 3.863078912 GB, Q5_0 2.856730624 GB, BF16 1.476395008 GB, Q6_K 0.061501440 GB, and F32 0.046056568 GB. Fixed non-expert tensor spans total 2.520870008 GB. Routed expert weight tensors total 15.130165248 GB across 30 layers and 128 expert indexes, or 0.118204416 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, K=V full-attention geometry, and separate sliding-layer K/V projections." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, HF CLI dry-run, model card metadata, base API and gated-config access checks, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected HB16 Q4_K_M artifact." }, "notes": "Use this profile for the API-selected ReadyArt Melody1437 HB16 Q4_K_M GGUF text artifact. Do not infer lower-bit, imatrix, or non-HB16 sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "readyart--serenity-26b-a4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ReadyArt/Serenity-26B-A4B-GGUF", "title": "ReadyArt Serenity 26B A4B GGUF HB16 Q6_K", "summary": "Audited memory-side text-decode bounds profile for the API-selected HB16 Q6_K GGUF artifact of ReadyArt Serenity 26B A4B.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "ReadyArt/Serenity-26B-A4B", "relation": "quantized", "source": "Hugging Face model card/API metadata, gated base access check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of ReadyArt/Serenity-26B-A4B, which itself records google/gemma-4-26B-A4B-it as its base model. The ReadyArt base config is manually gated in this audit environment, so this profile uses the public selected GGUF header as the architecture source instead of claiming a direct base-config comparison." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 23.509242304, "main_resident_weight_gb": 23.493419456, "auxiliary_resident_weight_gb": 0.015822848, "fixed_weight_gb": 2.915443136, "routed_expert_weight_gb": 0.16076544, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for Serenity-26B-A4B-HB16-Q6_K.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors", "shared_expert_notes": "The GGUF header records 8 active / 128 total experts. Dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight are always-on/shared tensors and are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects Serenity-26B-A4B-HB16-Q6_K.gguf. A GGUF v3 range-read found 658 tensors and 52 metadata entries. Tensor spans total 23.493419456 GB, while the linked file is 23.509242304 GB. Routed expert weight tensors total 20.577976320 GB across 30 layers and 128 expert indexes, or 0.160765440 GB per expert index. Non-expert tensor spans total 2.915443136 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The GGUF sliding-window pattern marks layers 5, 11, 17, 23, and 29 as full attention. Gemma 4 full-attention layers use K=V behavior, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format." }, "notes": "This profile targets the API-selected main text GGUF artifact. Other GGUF siblings are separate artifacts and should get separate profiles if selected by a workload." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.9316811303619132, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-hb16-q6-k-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, image handling, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The selected artifact is the HB16 Q6_K GGUF because HF API gguf.totalFileSize exactly matches Serenity-26B-A4B-HB16-Q6_K.gguf. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "ReadyArt Serenity 26B A4B GGUF API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Serenity-26B-A4B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "At commit 644908e4cee7807801cc6363cfd5cc4c7240b19a, the live API records a public non-gated GGUF repo with base_model ReadyArt/Serenity-26B-A4B, Apache-2.0 license metadata, region:us, 151391 downloads, GGUF architecture gemma4, context_length 262144, gguf.total 25233142046, and gguf.totalFileSize 23509242304." }, { "label": "ReadyArt Serenity 26B A4B GGUF model card", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B-GGUF/raw/644908e4cee7807801cc6363cfd5cc4c7240b19a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "runtime_format" ], "notes": "The card records base_model ReadyArt/Serenity-26B-A4B, base_model_relation quantized, Apache-2.0 licensing, Gemma 4 roleplay/instruct tags, and text-layer LoRA training over the Serenity base." }, { "label": "ReadyArt Serenity 26B A4B base API metadata and gated config check", "url": "https://huggingface.co/api/models/ReadyArt/Serenity-26B-A4B", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base API metadata records a public but manually gated safetensors repo, base_model google/gemma-4-26B-A4B-it, Apache-2.0 license metadata, and safetensors total 26544131376 parameters. Raw base config access returned HTTP 401 in this audit environment, so this profile does not claim direct base-config compatibility." }, { "label": "ReadyArt Serenity selected GGUF linked-object HEAD checks", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B-GGUF/tree/644908e4cee7807801cc6363cfd5cc4c7240b19a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Exact HEAD checks found Serenity-26B-A4B-HB16-Q6_K.gguf is 23.509242304 GB, exactly matching API gguf.totalFileSize. Nearby alternatives include HB16 Q8_0 at 27.551920064 GB, IQ4_XS at 14.063809472 GB, Q4_K_M at 16.796016576 GB, Q5_K_M at 19.132891072 GB, Q6_K at 22.638399936 GB, and Q8_0 at 26.859859904 GB." }, { "label": "ReadyArt Serenity HB16 Q6_K GGUF range-read tensor index", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B-GGUF/resolve/644908e4cee7807801cc6363cfd5cc4c7240b19a/Serenity-26B-A4B-HB16-Q6_K.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 658 tensors and 52 metadata entries. Tensor spans sum to 23.493419456 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015822848 GB. Tensor spans split into Q6_K 13.692940800 GB, Q8_0 8.278026240 GB, BF16 1.476395008 GB, and F32 0.046057408 GB. Fixed non-expert tensor spans total 2.915443136 GB. Routed expert weight tensors total 20.577976320 GB across 30 layers and 128 expert indexes, or 0.160765440 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, K=V full-attention geometry, and separate sliding-layer K/V projections." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card metadata, base API and gated-config access checks, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected HB16 Q6_K artifact." }, "notes": "Use this profile for the API-selected ReadyArt Serenity HB16 Q6_K GGUF text artifact. Do not infer lower-bit or non-HB16 sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "recviking--mistral-medium-3-5-128b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RecViking/Mistral-Medium-3.5-128B-NVFP4", "title": "RecViking Mistral Medium 3.5 128B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RecViking llm-compressor NVFP4 package of Mistral Medium 3.5 128B.", "model_family": "mistral3-dense-vlm", "base_model_proof": { "base_model": "mistralai/Mistral-Medium-3.5-128B", "relation": "quantized", "source": "Hugging Face model-card metadata, pinned served config, pinned base config comparison, recipe, and direct safetensors-header range reads", "config_compatible": true, "notes": "The card and API metadata identify this as an NVFP4 quantization of mistralai/Mistral-Medium-3.5-128B. Manual comparison against the pinned base config at ed0c85631cf20929c29bc36d7cb37519ed0cdd28 found no differences across the checked root, text_config, and vision_config geometry fields: Mistral3ForConditionalGeneration, 88 Ministral3 text layers, 96 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, no sliding window, untied embeddings, Pixtral vision tower, and multimodal projector geometry. The target adds compressed-tensors NVFP4 quantization metadata while preserving the checked base architecture." }, "architecture": { "canonical_architecture_id": "mistral-medium-3-5-128b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 127.704210176, "swept_params_b": 123.415375872, "auxiliary_resident_params_b": 4.288834304, "resident_weight_gb": 80.3172048, "swept_weight_gb": 71.739536192, "auxiliary_resident_weight_gb": 8.577668608, "resident_parameter_scope": "base logical Mistral Medium 3.5 128B parameters with exact stored compressed-tensors NVFP4/BF16/F8/F32 safetensors bytes", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens.weight, model.vision_tower, and model.multi_modal_projector are resident for the multimodal package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from both safetensors shard headers because the artifact mixes packed U8 NVFP4 payload tensors, F8_E4M3 scale tensors, tiny F32 scalar scales, and unquantized BF16 embeddings, lm_head, vision, projector, norms, and ignored modules. Logical parameter counts count each packed U8 element as two logical 4-bit weights and exclude scale tensors from logical model parameters, matching the 127.704210176B logical architecture." }, "kv_adapter": { "kind": "full_context", "layers": 88, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text config records sliding_window null and no recurrent or compressed-state attention path, so the ordinary text profile charges full-context K and V streams for all 88 Ministral3 decoder layers." }, "notes": "Mistral3ForConditionalGeneration is multimodal and includes a resident Pixtral vision tower plus multimodal projector. This profile models ordinary cached text decode after any multimodal prefill, not vision prefill throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "turboquant_4bit_nc", "kv_store_bytes_per_scalar": 0.5, "kv_read_format": "turboquant_4bit_nc", "kv_read_bytes_per_scalar": 0.5, "runtime_format": "vllm-compressed-tensors-nvfp4-turboquant-4bit-kv-mistral3-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored NVFP4/BF16/F8/F32 safetensors bytes and TurboQuant 4-bit KV bytes for the documented vLLM launch. NVFP4 dequantization, activation traffic, cache writes, TurboQuant implementation overhead, and compute are outside this memory-side bound.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights/activations and kv_cache_scheme null. The model card's tested long-context vLLM command explicitly uses --kv-cache-dtype turboquant_4bit_nc, so this profile represents that documented serving configuration rather than a generic BF16-KV launch." }, "evidence": [ { "label": "RecViking Mistral Medium 3.5 128B NVFP4 API metadata", "url": "https://huggingface.co/api/models/RecViking/Mistral-Medium-3.5-128B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA dc2c3c2962bf46c3a5335d19b1f4b4725011da32, the live API records a public non-gated safetensors repo with mistral3, vLLM, compressed-tensors, nvfp4, quantized, multimodal, base_model:mistralai/Mistral-Medium-3.5-128B, base_model:quantized:mistralai/Mistral-Medium-3.5-128B, license:other, 8-bit, and region:us tags. Current live downloads are 18825; the catalog row keeps the original qualifying scrape count of 106484 until the over-100k working set is regenerated. The API safetensors block reports F32 1232, BF16 5901622016, F8_E4M3 7612661760, U8 60901294080, and total 74415579088 storage-accounting elements." }, { "label": "RecViking Mistral Medium 3.5 128B NVFP4 model card", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/raw/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "quantized_module_scope", "kv_store_format" ], "notes": "The card describes an llm-compressor NVFP4 W4A4-FP4 quantization of mistralai/Mistral-Medium-3.5-128B in compressed-tensors nvfp4-pack-quantized format. It states the vision tower, multimodal projector, embeddings, and lm_head remain BF16 while the 88 Ministral3 decoder blocks are FP4. Its proven vLLM launch uses --kv-cache-dtype turboquant_4bit_nc and says long-context inference at 150k or more tokens was verified with TurboQuant 4-bit KV cache." }, { "label": "RecViking Mistral Medium 3.5 128B NVFP4 config", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/raw/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The pinned config records Mistral3ForConditionalGeneration, model_type mistral3, dtype bfloat16, compressed-tensors format nvfp4-pack-quantized, 340 ignored quantization entries covering vision and projector tensors plus lm_head, kv_cache_scheme null, text_config model_type ministral3, hidden size 12288, intermediate size 28672, 88 layers, 96 attention heads, 8 KV heads, head_dim 128, sliding_window null, 262144 max position embeddings, vocab size 131072, tie_word_embeddings false, and resident Pixtral vision_config with 48 layers." }, { "label": "RecViking Mistral Medium 3.5 128B NVFP4 recipe", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/raw/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/recipe.yaml", "source_type": "config", "supports": [ "serving", "quantization", "quantized_module_scope" ], "notes": "The recipe records a QuantizationModifier targeting Linear modules with scheme NVFP4 and ignore patterns for lm_head, vision_tower, and multi_modal_projector." }, { "label": "Mistral Medium 3.5 128B base config comparison", "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/raw/ed0c85631cf20929c29bc36d7cb37519ed0cdd28/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible" ], "notes": "Manual comparison found no differences across 22 checked profile-relevant root, text_config, and vision_config fields between the pinned base config and this NVFP4 artifact. The target adds quantization metadata while preserving the checked architecture geometry." }, { "label": "RecViking Mistral Medium 3.5 128B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/resolve/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 80317204800 bytes and total_parameters 74415579088 across two shards. Direct range reads of both shard headers found 3081 tensors totaling exactly 80.317204800 GB of tensor payload, with linked file bytes totaling 80.317612560 GB and 0.000407760 GB of safetensors header/container overhead outside tensor payloads. Payload dtype split is U8 60.901294080 GB, BF16 11.803244032 GB, F8_E4M3 7.612661760 GB, and F32 0.000004928 GB. The ordinary text swept subset, model.language_model excluding embed_tokens plus lm_head, totals 71.739536192 GB. Resident-only tensors total 8.577668608 GB: input embeddings 3.221225472 GB, vision tower 4.991404288 GB, and multimodal projector 0.365038848 GB. Header buckets are language MLP 52.319750208 GB, language self-attention 16.194210560 GB, lm_head 3.221225472 GB, input embedding 3.221225472 GB, vision tower 4.991404288 GB, projector 0.365038848 GB, layer norms 0.004325376 GB, and final norm 0.000024576 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current Hugging Face API metadata, pinned model card, pinned compressed-tensors NVFP4 config and recipe, pinned base config comparison, direct two-shard safetensors header range reads, and linked-object HEAD checks." }, "notes": "This profile supersedes the generated metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted BF16 embeddings, output head, vision/projector tensors, F8 scales, F32 scales, and the documented TurboQuant 4-bit KV serving path." }, { "id": "redhatai--deepseek-coder-v2-lite-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8", "title": "RedHatAI DeepSeek Coder V2 Lite Instruct FP8", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI static FP8 package of DeepSeek-Coder-V2-Lite-Instruct.", "model_family": "deepseek-coder-v2-lite-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "relation": "quantized", "source": "Hugging Face model card, served FP8 config, audited BF16 base config comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The model card identifies this as a quantized version of deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct. Manual config comparison against the audited BF16 base found the same DeepseekV2ForCausalLM architecture, layer count, hidden sizes, expert routing fields, attention dimensions, context length, and tied-embedding setting. The RedHatAI config adds static FP8 quantization metadata, _name_or_path, ep_size 1, and a newer Transformers version." }, "architecture": { "canonical_architecture_id": "deepseek-coder-v2-lite", "max_context_tokens": 163840, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.129490408, "main_resident_weight_gb": 15.710060008, "auxiliary_resident_weight_gb": 0.4194304, "fixed_weight_gb": 1.315168744, "routed_expert_weight_gb": 0.224920176, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary vLLM/Transformers text decode through layers 0-26, model.norm.weight, and lm_head.weight, excluding the resident-only input embedding matrix", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 2. Shared expert tensors are not under the routed experts.* tensor namespace and are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the static FP8 artifact mixes F8_E4M3 weights, BF16 embeddings/head tensors, and tiny F32 input_scale and weight_scale tensors. Routed expert tensors are byte-uniform across 64 expert indexes; routed_expert_weight_gb is the exact routed expert byte total divided by 64." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "The served config matches the audited BF16 base. The pinned custom DeepSeek attention path constructs cached key_states with 16 heads and q_head_dim 192, composed of 128 no-RoPE dimensions plus 64 RoPE dimensions." }, { "kind": "full_context", "layers": 27, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "The served config matches the audited BF16 base. The pinned custom DeepSeek attention path stores value_states with 16 heads and v_head_dim 128 before past_key_value.update." } ], "notes": "The architecture is MLA-style internally, but the referenced custom Transformers code expands compressed_kv into key_states and value_states before cache update. The RedHatAI FP8 config does not declare KV-cache quantization, so Bounds Engine v1 charges expanded BF16 K/V cache streams for all 27 decoder layers." }, "notes": "The served FP8 config records 27 ordinary hidden layers, with layer 0 dense and layers 1-26 MoE. There is no separate MTP/next-token-prediction layer in this checkpoint." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-autofp8-static-expanded-kv-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights, BF16 tensors, and F32 scale tensors from safetensors headers. FP8 dequantization, static activation scaling, activation traffic, routing overhead, and kernel efficiency remain outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and AutoFP8-style static activation FP8 quantization with lm_head ignored. It has quant_method fp8 but no kv_cache_scheme or vLLM command requesting FP8 KV, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI DeepSeek Coder V2 Lite FP8 API metadata", "url": "https://huggingface.co/api/models/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "weight_format", "dtype_split" ], "notes": "At repo SHA efe1ced428db63e7ccbcf367334596f77e9af140, the API records a public/non-gated text-generation repo with deepseek_v2, fp8, vLLM, custom_code, text-generation-inference, endpoints_compatible, and region:us tags. Current downloads are 245921. The API safetensors block reports BF16: 422964736, F8_E4M3: 15283519488, and total: 15706484224 tensor elements." }, { "label": "RedHatAI DeepSeek Coder V2 Lite FP8 model card", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "quantized_module_scope" ], "notes": "The card describes this as a quantized version of DeepSeek-Coder-V2-Lite-Instruct for vLLM >= 0.5.2. It states that weights and activations are quantized to FP8, only Linear operators inside transformer blocks are quantized, symmetric per-tensor quantization is used, lm_head is ignored, and expert gates are kept at original precision." }, { "label": "RedHatAI DeepSeek Coder V2 Lite FP8 config", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8/raw/efe1ced428db63e7ccbcf367334596f77e9af140/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records DeepseekV2ForCausalLM, model_type deepseek_v2, bfloat16 dtype, 27 hidden layers, first_k_dense_replace 1, hidden_size 2048, intermediate_size 10944, moe_intermediate_size 1408, 16 attention heads, 16 KV heads, kv_lora_rank 512, q_lora_rank null, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 64 routed experts, 6 experts per token, 2 shared experts, tie_word_embeddings false, 163840 max position embeddings, and static FP8 quantization with lm_head ignored." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct base config comparison", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison against the already audited BF16 base config found no profile-relevant geometry differences after excluding quantization_config, auto_map, dtype label, _name_or_path, ep_size, and Transformers version." }, { "label": "DeepSeek-Coder-V2-Lite-Instruct custom modeling file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/modeling_deepseek.py", "source_type": "manual_review", "supports": [ "kv_adapter", "moe_routing", "ordinary_decode_scope" ], "notes": "The RedHatAI config auto_map points to the official DeepSeek-Coder-V2-Lite-Instruct custom code. The already audited file builds layer 0 as dense MLP, layers 1-26 as DeepseekV2MoE with routed top-k experts plus shared_experts, and expands compressed_kv into key_states with 16 heads x 192 dims and value_states with 16 heads x 128 dims before past_key_value.update." }, { "label": "RedHatAI DeepSeek Coder V2 Lite FP8 safetensors index and headers", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8/resolve/efe1ced428db63e7ccbcf367334596f77e9af140/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index maps 15653 tensors across four safetensors shards and records total_size 16129490408 bytes. Direct range-read headers match exactly: tensor payloads total 16.129490408 GB, split into F8_E4M3 15.283519488 GB, BF16 0.845929472 GB, and F32 0.000041448 GB. model.embed_tokens.weight is 0.419430400 GB resident-only. Ordinary main tensors excluding input embeddings sum to 15.710060008 GB. Routed expert tensors under model.layers.1-26.mlp.experts.* sum to 14.394891264 GB, exactly 0.224920176 GB per expert index. Fixed ordinary-decode traffic including attention, dense layer 0 MLP, routers, shared experts, norms, lm_head, and scale tensors is 1.315168744 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, audited BF16 base config comparison, referenced custom modeling code, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as flat FP8 active metadata, undercounted active MoE traffic, and used a rounded full-KV coefficient. It is an ordinary text-decode profile for the RedHatAI static FP8 serving artifact, not a claim about runtime-specific latent MLA cache compression." }, { "id": "redhatai--diffusiongemma-26b-a4b-it-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic", "title": "RedHatAI DiffusionGemma 26B A4B IT FP8 Dynamic", "summary": "Unsupported profile stub with exact resident tensor evidence for the RedHatAI compressed-tensors FP8 dynamic DiffusionGemma serving artifact.", "model_family": "diffusion-gemma-block-diffusion-moe", "base_model_proof": { "base_model": "google/diffusiongemma-26B-A4B-it", "relation": "quantized", "source": "Hugging Face base_model metadata, model card, served config, recipe, and existing Google DiffusionGemma base profile", "config_compatible": true, "notes": "The served config preserves DiffusionGemmaForBlockDiffusion, 256-token canvas generation, 30 decoder layers, hybrid local/global attention, 128 experts, top_k_experts 8, and 262144 max position embeddings while adding compressed-tensors FP8 dynamic quantization metadata." }, "architecture": { "canonical_architecture_id": "diffusion-gemma-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 27.197665496, "main_resident_weight_gb": 26.052076604, "auxiliary_resident_weight_gb": 1.145588892, "fixed_weight_gb": 3.181651004, "routed_expert_weight_gb": 0.1786752, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f8_e4m3", "traffic_scope": "Exact decoder tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because DiffusionGemma block diffusion is not ordinary one-output-token autoregressive decode.", "auxiliary_scope": "model.encoder tensors are resident for the multimodal encoder/cache path but are not enough to define ordinary token-by-token swept traffic.", "shared_expert_notes": "The config records top_k_experts 8 and 128 routed experts. The checkpoint also stores dense decoder tensors outside model.decoder.layers.*.experts.*, so those always-on/shared tensors are included in fixed_weight_gb.", "notes": "Header-derived bytes are used. model.decoder tensors total 26.052076604 GB, model.encoder tensors total 1.145588892 GB, and no lm_head tensor is stored separately. Routed expert tensors total 22.870425600 GB and divide exactly into 128 uniform expert groups of 0.178675200 GB." }, "kv_adapter": { "kind": "unknown", "reason": "DiffusionGemma uses block diffusion over a 256-token canvas with a decoder that applies bidirectional attention over the generation canvas and then appends fully denoised canvases to cache. Bounds Engine v1 only models ordinary autoregressive per-output-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "The RedHatAI compressed-tensors config records FP8 weights and dynamic per-token FP8 activations but does not define an ordinary autoregressive KV traffic model. A production profile needs a dedicated block-diffusion adapter with canvas length, denoising iteration count, sampler behavior, canvas self-attention traffic, cross-attention/context-cache traffic, and block append policy." }, "notes": "This profile intentionally fails closed even though config, quantization metadata, and tensor headers are accessible, because the supported comparison math does not model DiffusionGemma's block-diffusion generation algorithm." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-block-diffusion-vllm-compressed-tensors-fp8-dynamic", "dequantization_notes": "The config and recipe record compressed-tensors FP8 weights with dynamic per-token FP8 activations. Bounds Engine v1 does not turn those bytes into production tok/s for this repo because the generation algorithm is block diffusion rather than ordinary autoregressive decode.", "notes": "The model card documents vLLM serving with llm-compressor/compressed-tensors FP8 dynamic quantization." }, "evidence": [ { "label": "RedHatAI DiffusionGemma FP8 Dynamic Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "weight_format", "total_params_b" ], "notes": "At commit 9133fd54d9610b569b17341ca7e24efb9687ef98, the API reports a public non-gated repo with base_model google/diffusiongemma-26B-A4B-it, fp8/vLLM/compressed-tensors/llm-compressor tags, 277204 downloads, region:us, and safetensors parameters BF16 1357117548, F8_E4M3 24483430400, total 25840547948." }, { "label": "RedHatAI DiffusionGemma FP8 Dynamic served config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic/raw/9133fd54d9610b569b17341ca7e24efb9687ef98/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "weight_format", "unsupported_reason" ], "notes": "The config records DiffusionGemmaForBlockDiffusion, model_type diffusion_gemma, canvas_length 256, 30 decoder layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 16 attention heads, 8 local KV heads, 2 global KV heads, 128 experts, top_k_experts 8, 262144 max position embeddings, tied embeddings, and compressed-tensors FP8 dynamic metadata with kv_cache_scheme null." }, { "label": "RedHatAI DiffusionGemma FP8 Dynamic model card and recipe", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "unsupported_reason", "runtime_format" ], "notes": "The card states this is an FP8 quantized version of google/diffusiongemma-26B-A4B-it with weights and activations quantized to FP8 using vLLM/llm-compressor in compressed-tensors format. The recipe targets Linear modules and ignores lm_head, embeddings, routers, vision tower, and self-conditioning modules." }, { "label": "RedHatAI DiffusionGemma FP8 Dynamic generation config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic/raw/9133fd54d9610b569b17341ca7e24efb9687ef98/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config records max_denoising_steps 48, max_new_tokens 256, EntropyBoundSamplerConfig, confidence_threshold 0.005, stability_threshold 1, t_min 0.4, and t_max 0.8, confirming a diffusion sampling policy rather than a normal one-token decode loop." }, { "label": "RedHatAI DiffusionGemma FP8 Dynamic safetensors index and header", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic/raw/9133fd54d9610b569b17341ca7e24efb9687ef98/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 27.197665496 GB in a single safetensors file. A direct range read found a 3199128-byte header and 24232 tensors: 2.714235096 GB BF16 and 24.483430400 GB F8_E4M3. model.decoder tensors total 26.052076604 GB; model.encoder tensors total 1.145588892 GB. model.decoder.embed_tokens.weight is 1.476395008 GB, and there is no separate lm_head tensor. Non-expert decoder tensors total 3.181651004 GB. Routed expert tensors under model.decoder.layers.*.experts.* total 22.870425600 GB and divide exactly into 128 uniform expert groups of 0.178675200 GB." }, { "label": "Google DiffusionGemma 26B A4B IT base profile", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it", "source_type": "manual_review", "supports": [ "unsupported_reason", "base_model_proof" ], "notes": "The base BF16 repo is already fail-closed in Local Frontier for the same architectural reason: Bounds Engine v1 lacks a block-diffusion throughput adapter." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from current HF API metadata, the pinned served config, recipe, generation config, model card, safetensors index, direct single-file header byte grouping, and the existing base DiffusionGemma profile. Marked unsupported because Bounds Engine v1 lacks a block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model block-diffusion generation over a denoised canvas, so ordinary autoregressive throughput would be misleading even though the RedHatAI compressed-tensors resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated DiffusionGemma block-diffusion adapter exists." }, { "id": "redhatai--diffusiongemma-26b-a4b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "RedHatAI/diffusiongemma-26B-A4B-it-NVFP4", "title": "RedHatAI DiffusionGemma 26B A4B IT NVFP4", "summary": "Unsupported profile stub with exact resident tensor evidence for the RedHatAI compressed-tensors NVFP4 DiffusionGemma serving artifact.", "model_family": "diffusion-gemma-block-diffusion-moe", "base_model_proof": { "base_model": "google/diffusiongemma-26B-A4B-it", "relation": "quantized", "source": "Hugging Face base_model metadata, model card, served config, recipe, and existing Google DiffusionGemma base profile", "config_compatible": true, "notes": "The served config preserves DiffusionGemmaForBlockDiffusion, 256-token canvas generation, 30 decoder layers, hybrid local/global attention, 128 experts, top_k_experts 8, and 262144 max position embeddings while adding compressed-tensors NVFP4 quantization metadata." }, "architecture": { "canonical_architecture_id": "diffusion-gemma-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 18.048607656, "main_resident_weight_gb": 16.903018764, "auxiliary_resident_weight_gb": 1.145588892, "fixed_weight_gb": 4.056559884, "routed_expert_weight_gb": 0.10036296, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f8_e4m3_u8_f32", "traffic_scope": "Exact decoder tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because DiffusionGemma block diffusion is not ordinary one-output-token autoregressive decode.", "auxiliary_scope": "model.encoder tensors are resident for the multimodal encoder/cache path but are not enough to define ordinary token-by-token swept traffic.", "shared_expert_notes": "The config records top_k_experts 8 and 128 routed experts. The checkpoint also stores dense decoder tensors outside model.decoder.layers.*.experts.*, so those always-on/shared tensors are included in fixed_weight_gb.", "notes": "Header-derived bytes are used. model.decoder tensors total 16.903018764 GB, model.encoder tensors total 1.145588892 GB, and no lm_head tensor is stored separately. Routed expert tensors total 12.846458880 GB and divide exactly into 128 uniform expert groups of 0.100362960 GB." }, "kv_adapter": { "kind": "unknown", "reason": "DiffusionGemma uses block diffusion over a 256-token canvas with a decoder that applies bidirectional attention over the generation canvas and then appends fully denoised canvases to cache. Bounds Engine v1 only models ordinary autoregressive per-output-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "The RedHatAI compressed-tensors config records NVFP4 weights and activations but does not define an ordinary autoregressive KV traffic model. A production profile needs a dedicated block-diffusion adapter with canvas length, denoising iteration count, sampler behavior, canvas self-attention traffic, cross-attention/context-cache traffic, and block append policy." }, "notes": "This profile intentionally fails closed even though config, quantization metadata, and tensor headers are accessible, because the supported comparison math does not model DiffusionGemma's block-diffusion generation algorithm." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 1, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 1, "runtime_format": "unsupported-block-diffusion-vllm-compressed-tensors-nvfp4", "dequantization_notes": "The config and recipe record compressed-tensors NVFP4 weights and activations. Bounds Engine v1 does not turn those bytes into production tok/s for this repo because the generation algorithm is block diffusion rather than ordinary autoregressive decode.", "notes": "The model card documents vLLM serving with llm-compressor/compressed-tensors NVFP4 quantization." }, "evidence": [ { "label": "RedHatAI DiffusionGemma NVFP4 Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "weight_format", "total_params_b" ], "notes": "At commit 8484b309d29a474b098ad57c47a9045e1fc4b48a, the API reports a public non-gated repo with base_model google/diffusiongemma-26B-A4B-it, nvfp4/vLLM/compressed-tensors/llm-compressor tags, 539635 downloads, region:us, and safetensors parameters BF16 2450530668, F8_E4M3 1460828160, U8 11686625280, F32 11623, total 15597995731." }, { "label": "RedHatAI DiffusionGemma NVFP4 served config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4/raw/8484b309d29a474b098ad57c47a9045e1fc4b48a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "weight_format", "unsupported_reason" ], "notes": "The config records DiffusionGemmaForBlockDiffusion, model_type diffusion_gemma, canvas_length 256, 30 decoder layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, 16 attention heads, 8 local KV heads, 2 global KV heads, 128 experts, top_k_experts 8, 262144 max position embeddings, tied embeddings, and compressed-tensors nvfp4-pack-quantized metadata." }, { "label": "RedHatAI DiffusionGemma NVFP4 model card and recipe", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "unsupported_reason", "runtime_format" ], "notes": "The card states this is an NVFP4 quantized version of google/diffusiongemma-26B-A4B-it with weights and activations quantized to NVFP4 using vLLM/llm-compressor in compressed-tensors format. The recipe targets Linear modules and ignores embeddings, self-attention, routers, vision tower, self-conditioning, and lm_head-style modules." }, { "label": "RedHatAI DiffusionGemma NVFP4 generation config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4/raw/8484b309d29a474b098ad57c47a9045e1fc4b48a/generation_config.json", "source_type": "config", "supports": [ "unsupported_reason", "generation_algorithm" ], "notes": "The generation config records max_denoising_steps 48, max_new_tokens 256, EntropyBoundSamplerConfig, confidence_threshold 0.005, stability_threshold 1, t_min 0.4, and t_max 0.8, confirming a diffusion sampling policy rather than a normal one-token decode loop." }, { "label": "RedHatAI DiffusionGemma NVFP4 safetensors index and header", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4/raw/8484b309d29a474b098ad57c47a9045e1fc4b48a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The index records total_size 18.048607656 GB in a single safetensors file. A direct range read found a 6090264-byte header and 47337 tensors: 4.901061336 GB BF16, 1.460828160 GB F8_E4M3, 11.686625280 GB U8, and 0.000092880 GB F32. model.decoder tensors total 16.903018764 GB; model.encoder tensors total 1.145588892 GB. model.decoder.embed_tokens.weight is 1.476395008 GB, and there is no separate lm_head tensor. Non-expert decoder tensors total 4.056559884 GB. Routed expert tensors under model.decoder.layers.*.experts.* total 12.846458880 GB and divide exactly into 128 uniform expert groups of 0.100362960 GB." }, { "label": "Google DiffusionGemma 26B A4B IT base profile", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it", "source_type": "manual_review", "supports": [ "unsupported_reason", "base_model_proof" ], "notes": "The base BF16 repo is already fail-closed in Local Frontier for the same architectural reason: Bounds Engine v1 lacks a block-diffusion throughput adapter." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Reviewed from current HF API metadata, the pinned served config, recipe, generation config, model card, safetensors index, direct single-file header byte grouping, and the existing base DiffusionGemma profile. Marked unsupported because Bounds Engine v1 lacks a block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model block-diffusion generation over a denoised canvas, so ordinary autoregressive throughput would be misleading even though the RedHatAI compressed-tensors resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated DiffusionGemma block-diffusion adapter exists." }, { "id": "redhatai--gemma-3-27b-it-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-3-27b-it-FP8-dynamic", "title": "RedHatAI Gemma 3 27B IT FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI FP8 dynamic Gemma 3 27B IT serving artifact.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-27b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata and the served RedHatAI config", "config_compatible": false, "notes": "The model card/API metadata identify google/gemma-3-27b-it as the base model, but that base repo remains gated in this audit environment. This profile therefore uses the public RedHatAI served config and safetensors headers directly instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.436247664, "swept_params_b": 27.013187328, "auxiliary_resident_params_b": 0.423060336, "resident_weight_gb": 29.274133728, "swept_weight_gb": 28.428013056, "auxiliary_resident_weight_gb": 0.846120672, "resident_parameter_scope": "safetensors_header_stored_f8_e4m3_bf16", "swept_parameter_scope": "ordinary text decode charges language_model tensors, including the tied language_model.model.embed_tokens.weight output projection", "auxiliary_scope": "vision_tower and multi_modal_projector tensors are resident for the multimodal package but not swept for ordinary text decode", "notes": "Range-read safetensors headers record 1681 tensors totaling 27436247664 stored elements and 29.274133728 GB payload bytes. The checkpoint has no separate language_model.lm_head.weight; language_model.model.embed_tokens.weight is BF16 [262208, 5376] and remains in swept ordinary decode traffic as the tied output projection. Resident-only non-text tensors are the BF16 vision tower plus multimodal projector." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window_pattern 6 over 62 language layers. Using the documented Gemma 3 pattern of five local layers followed by one global layer gives 10 full-context global layers." }, { "kind": "sliding_window", "layers": 52, "kv_heads": 16, "head_dim": 128, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 52 language layers use the config's 1024-token local sliding-window attention." } ], "notes": "Layered KV models ordinary text decode after any image prefill. The config records cache_implementation hybrid and kv_cache_scheme null, so the profile charges BF16 K and V streams." }, "notes": "Gemma3ForConditionalGeneration is multimodal. This profile models ordinary text decode, not vision encoder or image-prefill throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0669875154397133, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads and BF16 embeddings, vision tensors, norms, and quantization scale tensors from safetensors headers. Dynamic FP8 activation quantization, dequantization, compute throughput, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The config records compressed-tensors float quantization with FP8 weights, dynamic per-token FP8 activation quantization, bfloat16 model dtype, and kv_cache_scheme null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Gemma 3 27B FP8 Dynamic model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-3-27b-it-FP8-dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 306078afe860ef87821018f28078ecf762f31455, the API reports a public Apache-2.0 image-text-to-text repo with base_model google/gemma-3-27b-it, FP8/compressed-tensors/vLLM tags, region:us, 520563 downloads, and safetensors parameters split across BF16: 1837886064 and F8_E4M3: 25598361600 elements. The model card identifies the artifact as a quantized derivative of google/gemma-3-27b-it." }, { "label": "RedHatAI Gemma 3 27B FP8 Dynamic config", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic/raw/306078afe860ef87821018f28078ecf762f31455/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "sliding_window_pattern", "max_context_tokens", "serving", "vision_geometry" ], "notes": "The config records Gemma3ForConditionalGeneration, gemma3_text, bfloat16 model dtype, compressed-tensors float quantization, FP8 weight quantization with per-channel static scaling, dynamic per-token FP8 activation quantization, kv_cache_scheme null, cache_implementation hybrid, 62 text layers, 5376 hidden size, 21504 intermediate size, 32 attention heads, 16 KV heads, 128 head dimension, 131072 max positions, 1024-token sliding window, sliding_window_pattern 6, and a 27-layer SigLIP vision tower." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes the repeating local/global attention pattern as five local attention layers followed by one global attention layer. Applied to the 62 layers in the served config, that yields 10 global layers and 52 local layers." }, { "label": "RedHatAI Gemma 3 27B FP8 Dynamic safetensors index", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic/raw/306078afe860ef87821018f28078ecf762f31455/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "total_params_b", "weight_format" ], "notes": "The safetensors index records total_size 29274133728 bytes across six shards. Direct header reads found tensor payload bytes matching that index total." }, { "label": "RedHatAI Gemma 3 27B FP8 Dynamic safetensors headers", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic/tree/306078afe860ef87821018f28078ecf762f31455", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_scope", "weight_format" ], "notes": "Range-reads of all six safetensors shard headers found 1681 tensors totaling 29.274133728 GB: F8_E4M3 tensors total 25.598361600 GB and BF16 tensors total 3.675772128 GB. language_model tensors total 28.428013056 GB; language_model.model.layers total 25.608741888 GB, language_model.model.norm.weight totals 0.000010752 GB, and language_model.model.embed_tokens.weight is BF16 [262208, 5376] totaling 2.819260416 GB. There is no language_model.lm_head.weight tensor. Resident-only vision_tower tensors total 0.833732064 GB and multi_modal_projector tensors total 0.012388608 GB." }, { "label": "Google Gemma 3 27B IT gated base access note", "url": "https://huggingface.co/google/gemma-3-27b-it", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base repo remains gated in this audit environment, matching the existing google/gemma-3-27b-it unsupported profile. This audited RedHatAI profile does not copy base geometry; it relies on the public derivative config and tensor headers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card metadata, served config, safetensors index, direct safetensors shard header range reads, and Gemma 3 local/global attention documentation." }, "notes": "This profile supersedes the generated row's full-context KV estimate with a Gemma 3 layered local/global KV adapter and exact mixed FP8/BF16 tensor payload bytes." }, { "id": "redhatai--gemma-3-27b-it-quantized-w4a16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-3-27b-it-quantized.w4a16", "title": "RedHatAI Gemma 3 27B IT W4A16", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI compressed-tensors W4A16 Gemma 3 27B IT serving artifact.", "model_family": "gemma3-dense", "base_model_proof": { "base_model": "google/gemma-3-27b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata and the served RedHatAI config", "config_compatible": false, "notes": "The model card/API metadata identify google/gemma-3-27b-it as the base model, but that base repo remains gated in this audit environment. This profile therefore uses the public RedHatAI served config and safetensors headers directly instead of copying or inferring from the gated base config." }, "architecture": { "canonical_architecture_id": "gemma-3-27b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 29.042024916, "swept_params_b": 27.209334372, "auxiliary_resident_params_b": 1.832690544, "resident_weight_gb": 19.68651264, "swept_weight_gb": 16.021131552, "auxiliary_resident_weight_gb": 3.665381088, "resident_parameter_scope": "safetensors_header_logical_int4_bf16_i64", "swept_parameter_scope": "ordinary text decode charges quantized language layer tensors, language norm, and the separate BF16 language_model.lm_head.weight output projection", "auxiliary_scope": "language_model.model.embed_tokens.weight input embedding, vision_tower, and multi_modal_projector tensors are resident for the package but not swept as full matrices for each generated text token", "notes": "Range-read safetensors headers record 2116 tensors totaling 29.042024916B logical parameters and 19.686512640 GB payload bytes. The checkpoint stores packed I32 int4 tensors, BF16 unquantized tensors, and tiny I64 side tensors. Unlike the FP8 dynamic sibling, this artifact has both language_model.model.embed_tokens.weight and language_model.lm_head.weight as separate BF16 [262208, 5376] tensors. The input embedding is resident-only for ordinary decode; lm_head.weight remains in swept output-projection traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 16, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records sliding_window_pattern 6 over 62 language layers. Using the documented Gemma 3 pattern of five local layers followed by one global layer gives 10 full-context global layers." }, { "kind": "sliding_window", "layers": 52, "kv_heads": 16, "head_dim": 128, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 52 language layers use the config's 1024-token local sliding-window attention." } ], "notes": "Layered KV models ordinary text decode after any image prefill. The config records cache_implementation hybrid and kv_cache_scheme null, so the profile charges BF16 K and V streams." }, "notes": "Gemma3ForConditionalGeneration is multimodal. This profile models ordinary text decode, not vision encoder or image-prefill throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6778629485010252, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-gptq-w4a16-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed int4 I32 tensors, BF16 ignored modules and scale tensors, and I64 side tensors from safetensors headers. GPTQ dequantization, activation traffic, compute throughput, vision encoder throughput, and scheduler behavior are outside Bounds Engine v1.", "notes": "The model card describes INT4 weight quantization and FP16 activation quantization. The served config records compressed-tensors pack-quantized int4 weights with group size 128, bfloat16 model dtype, and kv_cache_scheme null; this profile therefore charges exact stored tensor bytes for weights and BF16 KV cache traffic." }, "evidence": [ { "label": "RedHatAI Gemma 3 27B W4A16 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-3-27b-it-quantized.w4a16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format", "total_params_b" ], "notes": "At commit 2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9, the live API reports a public non-gated Gemma-licensed image-text-to-text repo with base_model google/gemma-3-27b-it, vLLM/vision/w4a16/compressed-tensors tags, region:us, 323684 downloads, and safetensors logical parameters I64: 868, I32: 25598361600, BF16: 3443662448, total: 29042024916. The model card identifies the artifact as a quantized derivative of google/gemma-3-27b-it, with INT4 weight quantization and FP16 activation quantization." }, { "label": "RedHatAI Gemma 3 27B W4A16 config", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "sliding_window_pattern", "max_context_tokens", "serving", "vision_geometry" ], "notes": "The config records Gemma3ForConditionalGeneration, gemma3_text, bfloat16 model dtype, compressed-tensors pack-quantized int4 weights, group_size 128, weight actorder, kv_cache_scheme null, cache_implementation hybrid, 62 text layers, 5376 hidden size, 21504 intermediate size, 32 attention heads, 16 KV heads, 128 head dimension, 131072 max positions, 1024-token sliding window, sliding_window_pattern 6, and a 27-layer SigLIP vision tower." }, { "label": "RedHatAI Gemma 3 27B W4A16 quantization recipe", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope", "swept_weight_gb" ], "notes": "The llm-compressor GPTQ recipe targets Linear modules with 4-bit grouped int weights, group size 128, and ignores lm_head, embed_tokens, vision_tower, and multi_modal_projector modules. Safetensors headers show lm_head.weight is stored separately as BF16 and is charged as swept output-projection traffic for ordinary text decode." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes the repeating local/global attention pattern as five local attention layers followed by one global attention layer, and says larger Gemma 3 models support 128k context. Applied to the 62 layers in the served config, that yields 10 global layers and 52 local layers." }, { "label": "RedHatAI Gemma 3 27B W4A16 safetensors index", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "total_params_b", "weight_format" ], "notes": "The safetensors index records total_size 19686512640 bytes across four shards. Direct header reads found tensor payload bytes matching that index total." }, { "label": "RedHatAI Gemma 3 27B W4A16 safetensors headers", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/tree/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "vision_scope", "weight_format" ], "notes": "Range-reads of all four safetensors shard headers found 2116 tensors totaling 19.686512640 GB: I32 packed tensors total 12.799180800 GB, BF16 tensors total 6.887324896 GB, and I64 side tensors total 0.000006944 GB. Logical parameters total 29.042024916B after expanding weight_packed I32 tensors by 8. Ordinary swept text traffic totals 27.209334372B logical parameters / 16.021131552 GB: language MLP 11.087293344 GB, language self-attention 2.111900544 GB, other language layer tensors 0.002666496 GB, language norm 0.000010752 GB, and separate lm_head.weight 2.819260416 GB. Resident-only tensors total 1.832690544B logical parameters / 3.665381088 GB: language_model.model.embed_tokens.weight 2.819260416 GB, vision_tower 0.833732064 GB, and multi_modal_projector 0.012388608 GB." }, { "label": "Google Gemma 3 27B IT gated base access note", "url": "https://huggingface.co/google/gemma-3-27b-it", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The base repo remains gated in this audit environment, matching the existing google/gemma-3-27b-it unsupported profile. This audited RedHatAI profile does not copy base geometry; it relies on the public derivative config and tensor headers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card metadata, served config, quantization recipe, safetensors index, direct safetensors shard header range reads, and Gemma 3 local/global attention documentation." }, "notes": "This profile supersedes the generated row's incorrect I32 full-precision weight estimate with exact compressed-tensors W4A16 stored bytes and a Gemma 3 layered local/global KV adapter." }, { "id": "redhatai--gemma-4-12b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-12B-it-NVFP4", "title": "RedHatAI Gemma 4 12B IT NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI NVFP4 Gemma 4 12B Unified IT serving artifact.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, quantization recipe, and safetensors header audit", "config_compatible": true, "notes": "Manual comparison found matching Gemma4UnifiedForConditionalGeneration text, audio, vision, context, and attention geometry between the RedHatAI NVFP4 artifact and the Google BF16 IT repo. The RedHatAI artifact adds compressed-tensors NVFP4 quantization metadata and stores a separate BF16 lm_head.weight despite tie_word_embeddings true." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.1976368, "swept_params_b": 7.138623936, "auxiliary_resident_params_b": 1.059012864, "resident_weight_gb": 10.264054432, "swept_weight_gb": 8.146028704, "auxiliary_resident_weight_gb": 2.118025728, "resident_parameter_scope": "safetensors_header_stored_nvfp4_bf16_f8_e4m3_u8_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens, model.embed_audio, model.embed_vision, and model.vision_embedder tensors are resident for the package but not full-matrix swept for each generated text token", "notes": "Header-derived bytes are used because this compressed-tensors artifact stores NVFP4 packed tensors, FP8 scale tensors, F32 scalar scales, and unquantized BF16 tensors. The config records tie_word_embeddings true, but this artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection and treats model.language_model.embed_tokens.weight as a resident input table for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config records attention_k_eq_v true, num_global_key_value_heads 1, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Quantizing weights does not change the BF16 KV cache assumption because the config has kv_cache_scheme null." }, "notes": "Gemma4UnifiedForConditionalGeneration accepts text, image, video, and audio inputs without separate heavyweight encoders. This profile models ordinary text decode after any multimodal prefill, not input projection throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 1.2520747969707562, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-bf16-kv", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed NVFP4 payload tensors, FP8 scale tensors, F32 scalar scales, ignored BF16 modules, and BF16 output/input projection tensors from safetensors headers. Local FP4 activation quantization, dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights, group size 16, local FP4 activation quantization, bfloat16 model dtype, and kv_cache_scheme null. Exact resident and swept traffic use the header-derived GB fields above." }, "evidence": [ { "label": "RedHatAI Gemma 4 12B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-12B-it-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "weight_format", "total_params_b" ], "notes": "The live HF CLI/API response records repo SHA a1d2478a9b99cc444bc9f64839609d3a82ca9195, public ungated status, region:us tag, base_model google/gemma-4-12B-it, nvfp4/vLLM/compressed-tensors/GPTQ tags, 260994 downloads, and safetensors storage elements F32: 656, BF16: 2066415664, F8_E4M3: 681246720, U8: 5449973760, total: 8197636800. The API does not currently expose a license or pipeline_tag for this repo." }, { "label": "RedHatAI Gemma 4 12B NVFP4 README", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4", "source_type": "model_card", "supports": [ "weight_format", "activation_format", "auxiliary_resident_scope" ], "notes": "The README identifies this as a preliminary NVFP4 quantized google/gemma-4-12B-it artifact for vLLM/llm-compressor. It states weights and activations are quantized to NVFP4, tested against vLLM nightly, and created with a GPTQModifier while ignoring lm_head, embed_vision, embed_audio, and vision_embedder modules." }, { "label": "RedHatAI Gemma 4 12B NVFP4 config", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/raw/a1d2478a9b99cc444bc9f64839609d3a82ca9195/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4UnifiedForConditionalGeneration, compressed-tensors nvfp4-pack-quantized weights, local FP4 activation quantization, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 262144 max position embeddings, 16 attention heads, 8 sliding KV heads, 1 global KV head, 256 sliding head dimension, 512 global head dimension, and lightweight audio/vision projection configs." }, { "label": "RedHatAI Gemma 4 12B NVFP4 recipe", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/raw/a1d2478a9b99cc444bc9f64839609d3a82ca9195/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope" ], "notes": "The llm-compressor recipe targets Linear modules with scheme NVFP4, block size 128, static actorder, and ignores lm_head plus embed_vision, embed_audio, and vision_embedder paths. The safetensors header confirms those ignored modules remain stored as BF16 tensors." }, { "label": "Google Gemma 4 12B IT base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no profile-relevant geometry differences between the Google BF16 IT repo and this RedHatAI NVFP4 artifact after excluding quantization metadata and repository packaging. The RedHatAI artifact adds compressed-tensors metadata and a stored lm_head.weight." }, { "label": "RedHatAI Gemma 4 12B NVFP4 safetensors header", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/resolve/a1d2478a9b99cc444bc9f64839609d3a82ca9195/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The linked object size is 10264269928 bytes, with a 215488-byte header and 10264054432 tensor bytes across 1662 tensors. Stored tensors sum to 8.197636800B storage elements / 10.264054432 GB: U8 5.449973760 GB, F8_E4M3 0.681246720 GB, BF16 4.132831328 GB, and F32 0.000002624 GB. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens.weight plus lm_head.weight, sum to 7.138623936B storage elements / 8.146028704 GB. Resident-only tensors, defined as model.language_model.embed_tokens.weight plus model.embed_audio, model.embed_vision, and model.vision_embedder tensors, sum to 1.059012864B storage elements / 2.118025728 GB. Both model.language_model.embed_tokens.weight and lm_head.weight are present as separate BF16 tensors." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from HF CLI/API metadata, README, pinned compressed-tensors config, recipe.yaml, current base config comparison, linked object metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident input embedding and multimodal tensors from per-token swept language/logit weights." }, { "id": "redhatai--gemma-4-26b-a4b-it-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic", "title": "RedHatAI Gemma 4 26B A4B IT FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI FP8 dynamic Gemma 4 26B A4B IT serving artifact.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, expert routing fields, and vision geometry between the RedHatAI FP8 artifact and the Google BF16 base repo." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 28.63837046, "main_resident_weight_gb": 27.492781628, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 3.14596102, "routed_expert_weight_gb": 0.1786752, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges one BF16 vocabulary projection, non-expert language tensors, and expected distinct routed expert tensors", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The base Gemma card describes 1 shared expert and the config records top_k_experts 8. The FP8 dynamic quantization recipe ignores router, embed, vision, and lm_head tensors; shared/always-on MLP tensors remain non-expert language traffic and are charged in fixed_weight_gb.", "notes": "Header-derived bytes are used because the artifact stores FP8 weights plus BF16 scale tensors and unquantized BF16 modules. The checkpoint stores separate BF16 model.language_model.embed_tokens.weight and lm_head.weight tensors even though config tie_word_embeddings is true; resident bytes include both stored tensors, while swept decode traffic charges one full BF16 vocabulary projection rather than double-counting both." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Image/audio/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, not vision/audio encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8 weights, BF16 scale tensors, and unquantized BF16 modules from the safetensors header. Dynamic activation quantization, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors float quantization with FP8 weights, dynamic per-token FP8 activation quantization, bfloat16 model dtype, and kv_cache_scheme null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Gemma 4 26B FP8 Dynamic model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 099b66d6530675cb09eb75767f491dc783f848ba, the API reports an Apache-2.0 image-text-to-text repo with base_model google/gemma-4-26B-A4B-it, FP8/compressed-tensors/vLLM tags, and safetensors parameters split across BF16: 2077470030 and F8_E4M3: 24483430400 tensors. The model card states FP8 weight quantization, dynamic FP8 activation quantization, and that vision tower, embedding, output head, and MoE router layers remain original precision." }, { "label": "RedHatAI Gemma 4 26B FP8 Dynamic config", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic/raw/099b66d6530675cb09eb75767f491dc783f848ba/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors float quantization, FP8 weight quantization with per-channel static scaling, dynamic per-token FP8 activation quantization, ignore patterns for routers, vision, embeddings, and lm_head, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 30 text layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 128 experts, 8 experts per token, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The base card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this FP8 dynamic artifact; the RedHatAI artifact adds quantization_config while preserving the base architecture." }, { "label": "RedHatAI Gemma 4 26B FP8 Dynamic safetensors header", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic/resolve/099b66d6530675cb09eb75767f491dc783f848ba/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts" ], "notes": "The repo stores a single 28.641731036 GB safetensors file. A range-read of the 3,360,568-byte header found 24,199 tensors with tensor payloads totaling 28.63837046 GB: 24.4834304 GB F8_E4M3 tensors and 4.15494006 GB BF16 tensors. model.language_model tensors total 26.01638662 GB, lm_head.weight totals 1.476395008 GB, and resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. Ordinary text fixed traffic charges one full vocabulary projection plus non-expert language tensors for 3.14596102 GB. Routed expert tensors total 22.8704256 GB and divide exactly into 128 uniform expert groups of 0.1786752 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the RedHatAI model card, served config, base config comparison, HF API metadata, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this FP8 Dynamic repo, including the catalog row's incorrect routed-experts-per-token value." }, { "id": "redhatai--gemma-4-26b-a4b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-26B-A4B-it-NVFP4", "title": "RedHatAI Gemma 4 26B A4B IT NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI NVFP4 Gemma 4 26B A4B IT serving artifact.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, expert routing fields, and vision geometry between the RedHatAI NVFP4 artifact and the Google BF16 base repo." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.417035012, "main_resident_weight_gb": 15.27144618, "auxiliary_resident_weight_gb": 1.145588832, "fixed_weight_gb": 2.4249873, "routed_expert_weight_gb": 0.10036296, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_bf16_f8_e4m3_u8_f32", "traffic_scope": "ordinary text decode excludes resident vision tensors and charges the tied BF16 vocabulary projection, non-expert language tensors, and expected distinct NVFP4 routed expert tensors", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary text decode", "shared_expert_notes": "The base Gemma card describes 1 shared expert and the config records top_k_experts 8. The RedHatAI compressed-tensors recipe ignores router, vision, embed_vision, and lm_head modules, while the language shared/always-on MLP tensors remain in the non-expert language bucket charged in fixed_weight_gb.", "notes": "Header-derived bytes are used because this artifact stores NVFP4 packed weights, FP8 scale tensors, F32 scalar scales, and unquantized BF16 tensors. The checkpoint has model.language_model.embed_tokens.weight and no separate lm_head.weight tensor; because config tie_word_embeddings is true, ordinary text decode charges that stored vocabulary matrix once as the output projection rather than treating it as resident-only." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. Image/audio/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any multimodal prefill, not vision/audio encoder throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-bf16-kv", "dequantization_notes": "The memory-side bound charges stored packed NVFP4 weights, FP8/F32 scale tensors, and unquantized BF16 modules from the safetensors header. Local FP4 activation quantization, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights and local FP4 activation quantization, but quantization_config.kv_cache_scheme is null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Gemma 4 26B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-26B-A4B-it-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit 24eef434bc1089b25f8866c687fd015c63546735, the API reports an Apache-2.0 image-text-to-text repo with base_model google/gemma-4-26B-A4B-it, fp4/vLLM/llm-compressor/compressed-tensors tags, 1090344 downloads, and safetensors storage split across BF16: 1322505806, F8_E4M3: 1530214400, U8: 12241715200, and F32: 23450 elements. The card describes an NVFP4 variant of gemma-4-26B-A4B-it." }, { "label": "RedHatAI Gemma 4 26B NVFP4 README", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4", "source_type": "model_card", "supports": [ "weight_format", "activation_format", "auxiliary_resident_scope" ], "notes": "The README states FP4 weight quantization and FP4 activation quantization using NVFP4 for vLLM, with group size 16 for weights, local per-group activation scaling, and only transformer-block Linear operators quantized; the vision tower, embedding, output head, and MoE router layers remain original precision." }, { "label": "RedHatAI Gemma 4 26B NVFP4 config", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4/raw/24eef434bc1089b25f8866c687fd015c63546735/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors nvfp4-pack-quantized weights, local FP4 activation quantization, ignore patterns for routers, vision, embed_vision, and lm_head, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 30 text layers, five full-attention layers, 25 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 128 experts, 8 experts per token, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B IT model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it", "source_type": "model_card", "supports": [ "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The base card states hybrid local/global attention, 1024-token sliding window, 256K context, 8 active / 128 total experts, and 1 shared expert." }, { "label": "Google Gemma 4 26B A4B IT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this NVFP4 artifact; the RedHatAI artifact adds quantization_config while preserving the base architecture." }, { "label": "RedHatAI Gemma 4 26B NVFP4 safetensors index and header", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4/raw/24eef434bc1089b25f8866c687fd015c63546735/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "tie_word_embeddings" ], "notes": "The index records total_size 16417035012 bytes in one model.safetensors file. A range-read of the 6465576-byte header found 47648 tensors with tensor payloads totaling 16.417035012 GB: 12.2417152 GB U8, 1.5302144 GB F8_E4M3, 2.645011612 GB BF16, and 0.0000938 GB F32. model.language_model tensors total 15.27144618 GB and resident-only vision tensors under model.vision_tower plus model.embed_vision total 1.145588832 GB. The header found model.language_model.embed_tokens.weight at 1.476395008 GB and no separate lm_head.weight. Ordinary text fixed traffic charges that tied vocabulary projection once plus non-expert language tensors for 2.4249873 GB. Routed expert tensors total 12.84645888 GB and divide exactly into 128 uniform expert groups of 0.10036296 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the RedHatAI model card, README, served config, base config comparison, HF API metadata, safetensors index, and direct single-file safetensors header byte grouping." }, "notes": "This profile supersedes the scraped metadata estimate for this NVFP4 repo, including the catalog row's incorrect routed-experts-per-token value and BF16 KV behavior." }, { "id": "redhatai--gemma-4-31b-it-fp8-block", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-31B-it-FP8-block", "title": "RedHatAI Gemma 4 31B IT FP8 Block", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI FP8-block Gemma 4 31B IT serving artifact.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, and vision geometry between the RedHatAI FP8 artifact and the Google BF16 base repo." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 31.274876396, "swept_params_b": 30.69913286, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 33.263025112, "swept_weight_gb": 32.11153804, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "model.language_model safetensors headers, including FP8 block weights, BF16 scales, norms, and tied embedding/output projection", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary generated text tokens", "notes": "The config records tie_word_embeddings true and the index has no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. Header-derived bytes are used because the artifact stores FP8 block weights plus BF16 scale tensors and unquantized BF16 modules." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-block-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8 block weights, BF16 scale tensors, and unquantized BF16 modules from safetensors headers. Activation quantization, dequantization, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors FP8 block weights with 128x128 blocks, dynamic FP8 activation quantization, bfloat16 model dtype, and kv_cache_scheme null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Gemma 4 31B FP8 Block model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-31B-it-FP8-block", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit e1c5d725be4d3349e85334dc108e901dd99521e4, the API reports an Apache-2.0 image-text-to-text repo with base_model google/gemma-4-31B-it, FP8/compressed-tensors/vLLM tags, and safetensors parameters split across BF16: 1988148716 and F8_E4M3: 29286727680 tensors. The model card states FP8 block weight quantization, dynamic FP8 activation quantization, 128x128 weight blocks, group_size 128 activations, and that vision tower, embedding, and output head layers remain original precision." }, { "label": "RedHatAI Gemma 4 31B FP8 Block config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-block/raw/e1c5d725be4d3349e85334dc108e901dd99521e4/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors float quantization, FP8_BLOCK targets Linear, ignore patterns for vision, lm_head, and embed_tokens, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in the audited text_config and vision_config geometry fields between the Google BF16 repo and this FP8-block artifact; the RedHatAI artifact adds quantization_config while preserving the base architecture." }, { "label": "RedHatAI Gemma 4 31B FP8 Block safetensors index and shard headers", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-block/raw/e1c5d725be4d3349e85334dc108e901dd99521e4/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "The index records total_size 33263025112 bytes across two shards. Range-read safetensors headers found 1598 tensors totaling 33.263025112 GB: 29.28672768 GB F8_E4M3 tensors and 3.976297432 GB BF16 tensors. Language tensors under model.language_model total 30.69913286 stored tensor params / 32.11153804 GB and are swept for ordinary text decode, including BF16 scale tensors and the tied embedding/output projection. Resident-only vision tensors under model.vision_tower plus model.embed_vision total 0.575743536 BF16 params / 1.151487072 GB. The index has no separate lm_head.weight." }, { "label": "Google Gemma 4 31B IT BF16 profile", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_pattern", "base_model_proof" ], "notes": "This profile reuses the audited Gemma 4 31B text-decode KV layout from the Google BF16 profile because the RedHatAI config preserves the same 60-layer hybrid full/sliding attention architecture and does not enable KV-cache quantization." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the RedHatAI model card, served config, base config comparison, HF API metadata, safetensors index, and direct shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal weights from per-token swept language weights and keeps KV cache BF16 because the quantization config has kv_cache_scheme null." }, { "id": "redhatai--gemma-4-31b-it-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-31B-it-FP8-Dynamic", "title": "RedHatAI Gemma 4 31B IT FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI FP8-dynamic Gemma 4 31B IT serving artifact.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, and direct safetensors header grouping", "config_compatible": true, "notes": "Manual comparison found no differences in the checked top-level, text_config, and vision_config geometry fields between the RedHatAI FP8-dynamic artifact and the Google BF16 base repo. The RedHatAI artifact adds compressed-tensors FP8 dynamic quantization metadata while preserving the base architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 31.277317996, "swept_params_b": 30.70157446, "auxiliary_resident_params_b": 0.575743536, "resident_weight_gb": 33.267908312, "swept_weight_gb": 32.11642124, "auxiliary_resident_weight_gb": 1.151487072, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16", "swept_parameter_scope": "model.language_model safetensors headers, including FP8 channel-quantized weights, BF16 scales/norms, and tied embedding/output projection", "auxiliary_scope": "model.vision_tower and model.embed_vision are resident for the multimodal package but not swept for ordinary generated text tokens", "notes": "The config records tie_word_embeddings true and the headers have no separate lm_head.weight, so model.language_model.embed_tokens.weight is the tied output projection and remains swept for ordinary text decode. Header-derived bytes are used because the artifact stores FP8 channel-quantized weights plus unquantized BF16 modules and scales." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Every sixth language layer is full attention. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The 50 sliding-window layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Image/video prefill throughput is outside this text-decode profile. The quantization config has kv_cache_scheme null, so this profile keeps BF16 KV cache traffic." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored FP8 channel-quantized weights and unquantized BF16 tensors from safetensors headers. Dynamic per-token activation quantization, dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors FP8 static per-channel weights, dynamic per-token FP8 activations, bfloat16 model dtype, and kv_cache_scheme null, so this profile keeps KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Gemma 4 31B FP8 Dynamic model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-31B-it-FP8-dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "At commit c991bf46174b1523aa4208d602dfb072287719a7, the API reports an Apache-2.0 image-text-to-text repo with canonical id RedHatAI/gemma-4-31B-it-FP8-dynamic, base_model google/gemma-4-31B-it, FP8/compressed-tensors/vLLM tags, region:us, 329630 downloads, and safetensors parameters split across BF16: 1990590316 and F8_E4M3: 29286727680. The model card states FP8 weights and FP8 activation quantization using dynamic per-token activation quantization." }, { "label": "RedHatAI Gemma 4 31B FP8 Dynamic README", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic/raw/c991bf46174b1523aa4208d602dfb072287719a7/README.md", "source_type": "model_card", "supports": [ "weight_format", "activation_format", "auxiliary_resident_scope" ], "notes": "The README states FP8 weight quantization and FP8 activation quantization using dynamic per-token scaling for vLLM. It says weights are statically quantized with per-channel FP8 scaling, activations are dynamically quantized at inference time, and only transformer-block Linear operators are quantized; the vision tower, embedding, and output head remain original precision." }, { "label": "RedHatAI Gemma 4 31B FP8 Dynamic config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic/raw/c991bf46174b1523aa4208d602dfb072287719a7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors float quantization, FP8_DYNAMIC targets Linear, static per-channel FP8 weights, dynamic per-token FP8 activations, ignore patterns for vision, lm_head, and embed_tokens, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no differences in the checked top-level, text_config, and vision_config geometry fields between the Google BF16 repo and this RedHatAI FP8-dynamic artifact; the RedHatAI artifact adds quantization_config while preserving the base architecture." }, { "label": "RedHatAI Gemma 4 31B FP8 Dynamic safetensors index and shard headers", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic/raw/c991bf46174b1523aa4208d602dfb072287719a7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "Range-read safetensors headers found two shards with 1598 tensors totaling 33.267908312 GB: 29.286727680 GB F8_E4M3 tensors and 3.981180632 GB BF16 tensors. Stored tensor element count is 31.277317996B. Language tensors under model.language_model, including model.language_model.embed_tokens.weight as the tied output projection, total 30.701574460B stored elements / 32.116421240 GB and are swept for ordinary text decode. Resident-only vision tensors under model.vision_tower plus model.embed_vision total 0.575743536 BF16 params / 1.151487072 GB. The headers have no separate lm_head.weight." }, { "label": "Google Gemma 4 31B IT BF16 profile", "url": "https://huggingface.co/google/gemma-4-31B-it", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_pattern", "base_model_proof" ], "notes": "This profile reuses the audited Gemma 4 31B text-decode KV layout from the Google BF16 profile because the RedHatAI config preserves the same 60-layer hybrid full/sliding attention architecture and does not enable KV-cache quantization." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, README, pinned compressed-tensors config, current base config comparison, safetensors index, linked object metadata, and direct safetensors shard header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident multimodal weights from per-token swept language weights and keeps KV cache BF16 because the quantization config has kv_cache_scheme null." }, { "id": "redhatai--gemma-4-31b-it-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/gemma-4-31B-it-NVFP4", "title": "RedHatAI Gemma 4 31B IT NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI NVFP4 Gemma 4 31B IT serving artifact.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "Manual comparison found matching Gemma4ForConditionalGeneration shape, text layer geometry, local/global attention pattern, context fields, tied embeddings, dense/MoE setting, and vision geometry between the RedHatAI NVFP4 artifact and the Google BF16 base repo." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 19.86943248, "swept_params_b": 17.8844028, "auxiliary_resident_params_b": 1.98502968, "resident_weight_gb": 23.26508228, "swept_weight_gb": 19.29502292, "auxiliary_resident_weight_gb": 3.97005936, "resident_parameter_scope": "safetensors_header_stored_nvfp4_bf16_f8_e4m3_u8_f32", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens input table, model.vision_tower, and model.embed_vision tensors are resident for the package but not full-matrix swept for each generated text token", "notes": "Header-derived bytes are used because this compressed-tensors artifact stores NVFP4 packed weights, FP8 scale tensors, F32 scalar scales, and unquantized BF16 tensors. The config records tie_word_embeddings true, but this artifact contains a separate BF16 lm_head.weight. This profile charges lm_head.weight as the swept output projection and treats model.language_model.embed_tokens.weight as a resident input table for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. The config records attention_k_eq_v true, num_global_key_value_heads 4, and global_head_dim 512; safetensors headers have no v_proj tensors for full-attention layers." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 1024-token window. The header records separate k_proj and v_proj tensors for sliding-attention layers." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Quantizing weights does not change the BF16 KV cache assumption because the config has kv_cache_scheme null." }, "notes": "Gemma4ForConditionalGeneration is multimodal. This profile models ordinary text decode after any image prefill, not vision encoder throughput." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-bf16-kv", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed NVFP4 payload tensors, FP8 scale tensors, F32 scalar scales, ignored BF16 modules, and BF16 output/input projection tensors from safetensors headers. Local FP4 activation quantization, dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights, group size 16, local FP4 activation quantization, bfloat16 model dtype, and kv_cache_scheme null. Exact resident and swept traffic use the header-derived GB fields above." }, "evidence": [ { "label": "RedHatAI Gemma 4 31B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-31B-it-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b" ], "notes": "The live HF API response records repo SHA 34484c941404ed94fb46f04d7e0b04ac592b49a8, Apache-2.0 licensing, image-text-to-text pipeline, region:us tag, base_model google/gemma-4-31B-it, fp4/vLLM/llm-compressor/compressed-tensors tags, 342927 downloads, and safetensors storage elements F32: 820, BF16: 3395647340, F8_E4M3: 1830420480, U8: 14643363840, total: 19869432480. The card describes an NVFP4 variant of gemma-4-31B-it." }, { "label": "RedHatAI Gemma 4 31B NVFP4 README", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4", "source_type": "model_card", "supports": [ "weight_format", "activation_format", "auxiliary_resident_scope" ], "notes": "The README states FP4 weight quantization and FP4 activation quantization using NVFP4 for vLLM, with group size 16 for weights, local per-group activation scaling, and only transformer-block Linear operators quantized; the vision tower, embedding, and output head remain original precision." }, { "label": "RedHatAI Gemma 4 31B NVFP4 config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4/raw/34484c941404ed94fb46f04d7e0b04ac592b49a8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "serving", "auxiliary_resident_scope", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, compressed-tensors nvfp4-pack-quantized weights, local FP4 activation quantization, kv_cache_scheme null, tie_word_embeddings true, bfloat16 text config, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, 262144 max position embeddings, 32 attention heads, 16 sliding KV heads, 4 global KV heads, 256 sliding head dimension, 512 global head dimension, and Gemma 4 vision config." }, { "label": "RedHatAI Gemma 4 31B NVFP4 recipe", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4/raw/34484c941404ed94fb46f04d7e0b04ac592b49a8/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "auxiliary_resident_scope" ], "notes": "The llm-compressor recipe targets Linear modules with scheme NVFP4 and ignores vision, audio, lm_head, and embed modules, matching the mixed unquantized/packed safetensors header split." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "attention_pattern" ], "notes": "Manual comparison found no differences in 33 checked top-level, text_config, and vision_config geometry fields between the Google BF16 repo and this RedHatAI NVFP4 artifact; the RedHatAI artifact adds quantization_config while preserving the base architecture." }, { "label": "RedHatAI Gemma 4 31B NVFP4 safetensors header", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4/resolve/34484c941404ed94fb46f04d7e0b04ac592b49a8/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "tie_word_embeddings" ], "notes": "The single safetensors file was range-read directly. The linked object size is 23265401144 bytes, with a 318856-byte header and 23265082280 tensor bytes across 2419 tensors. Stored tensors sum to 19.869432480B storage elements / 23.265082280 GB: U8 14.643363840 GB, F8_E4M3 1.830420480 GB, BF16 6.791294680 GB, and F32 0.000003280 GB. Ordinary text swept tensors, defined as model.language_model excluding model.language_model.embed_tokens.weight plus lm_head.weight, sum to 17.884402800B storage elements / 19.295022920 GB. Resident-only tensors, defined as model.language_model.embed_tokens.weight plus model.vision_tower and model.embed_vision tensors, sum to 1.985029680B parameters / 3.970059360 GB. Both model.language_model.embed_tokens.weight and lm_head.weight are present." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, the model card, README, pinned compressed-tensors config, recipe.yaml, current base config comparison, linked object metadata, and direct safetensors header byte grouping." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident input embedding and multimodal tensors from per-token swept language/logit weights." }, { "id": "redhatai--llama-3-2-1b-instruct-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic", "title": "RedHatAI Llama 3.2 1B Instruct FP8 Dynamic", "summary": "Audited memory-side bounds profile for the RedHatAI compressed-tensors FP8 dynamic package of Llama 3.2 1B Instruct.", "model_family": "llama3.2-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-1B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata and served quantized config", "config_compatible": true, "notes": "The RedHatAI repo records meta-llama/Llama-3.2-1B-Instruct as its quantized base model. The base repo raw config is gated in this audit environment, but the public quantized config directly records the LlamaForCausalLM geometry used by this profile." }, "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.49885952, "swept_params_b": 1.236191232, "auxiliary_resident_params_b": 0.262668288, "resident_weight_gb": 2.024640512, "swept_weight_gb": 1.499303936, "auxiliary_resident_weight_gb": 0.525336576, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings true, but the safetensors file stores separate model.embed_tokens.weight and lm_head.weight tensors. This profile charges the stored resident artifact and keeps lm_head.weight in swept output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 16 layers, 8 KV heads, 64 head dimension, and 131072 max position embeddings." }, "notes": "Dense LlamaForCausalLM profile using the served RedHatAI FP8 dynamic config and exact stored safetensors bytes." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus BF16 embeddings, lm_head, norms, and per-channel weight_scale tensors from safetensors headers. Dynamic activation quantization and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and compressed-tensors float-quantized FP8 weights with dynamic token activation quantization, static channel weight quantization, and lm_head ignored. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Llama 3.2 1B Instruct FP8 Dynamic model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format" ], "notes": "At commit e23d444f8d7da0a3e556cae44a7d3c46f127e642, the API records a llama3.2-licensed text-generation repo with base_model meta-llama/Llama-3.2-1B-Instruct and tags fp8, vllm, and compressed-tensors. The API safetensors block reports tensor elements split across BF16: 525780992 and F8_E4M3: 973078528, total 1498859520." }, { "label": "RedHatAI Llama 3.2 1B Instruct FP8 Dynamic config", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic/raw/e23d444f8d7da0a3e556cae44a7d3c46f127e642/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, bfloat16, hidden size 2048, intermediate size 8192, 16 layers, 32 attention heads, 8 KV heads, 64 head dimension, 131072 max position embeddings, rope_theta 500000, vocab size 128256, tie_word_embeddings true, and compressed-tensors FP8 dynamic activation quantization with lm_head ignored." }, { "label": "RedHatAI Llama 3.2 1B Instruct FP8 Dynamic safetensors header", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic/resolve/e23d444f8d7da0a3e556cae44a7d3c46f127e642/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "Range-reading the single safetensors header found 259 tensors totaling 2.024640512 GB: 0.973078528 GB F8_E4M3 tensors and 1.051561984 GB BF16 tensors. model.embed_tokens.weight is BF16 with shape [128256, 2048] and contributes 262668288 tensor elements / 0.525336576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors plus model.norm.weight plus lm_head.weight total 1.499303936 GB swept traffic." }, { "label": "Meta Llama 3.2 1B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The base API metadata identifies the gated Meta repo as the public base model. Raw config access for the base repo is gated in this audit environment, so the quantized repo's public config is the direct geometry evidence for this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from RedHatAI API metadata, served FP8 dynamic config, safetensors header byte grouping, and Meta base API metadata." }, "notes": "This profile supersedes the scraped metadata estimate, which treated all stored weights as a flat 1 byte per parameter and missed BF16 resident/swept tensors stored by the compressed-tensors package." }, { "id": "redhatai--llama-3-2-1b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Llama-3.2-1B-Instruct-FP8", "title": "RedHatAI Llama 3.2 1B Instruct FP8", "summary": "Audited memory-side bounds profile for the RedHatAI compressed-tensors static FP8 package of Llama 3.2 1B Instruct.", "model_family": "llama3.2-dense", "base_model_proof": { "base_model": "meta-llama/Llama-3.2-1B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served quantized config, quantization recipe, and safetensors header", "config_compatible": true, "notes": "The RedHatAI repo records meta-llama/Llama-3.2-1B-Instruct as its quantized base model. The base repo raw config is gated in this audit environment, but the public quantized config directly records the LlamaForCausalLM geometry used by this profile." }, "architecture": { "canonical_architecture_id": "llama-3-2-1b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.498482912, "swept_params_b": 1.235814624, "auxiliary_resident_params_b": 0.262668288, "resident_weight_gb": 2.023887296, "swept_weight_gb": 1.49855072, "auxiliary_resident_weight_gb": 0.525336576, "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings true, but the safetensors file stores separate model.embed_tokens.weight and lm_head.weight tensors. This static FP8 artifact also stores per-tensor input_scale and weight_scale tensors for quantized Linear modules. The tiny scale tensors are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 16, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 16 layers, 8 KV heads, 64 head dimension, and 131072 max position embeddings." }, "notes": "Dense LlamaForCausalLM profile using the served RedHatAI static FP8 config and exact stored safetensors bytes." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-static-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus BF16 embeddings, lm_head, norms, input_scale tensors, and weight_scale tensors from safetensors headers. Static activation quantization, FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and compressed-tensors float-quantized FP8 weights with static tensor activation quantization, static tensor weight quantization, and lm_head ignored. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Llama 3.2 1B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Llama-3.2-1B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format" ], "notes": "At commit fb49430ca7e61099fa1ff30f12fecb290a8ebb65, the API records a public llama3.2-licensed text-generation repo with base_model meta-llama/Llama-3.2-1B-Instruct, compressed-tensors, region:us, and 884137 downloads. The API safetensors block reports tensor elements split across BF16: 525404384 and F8_E4M3: 973078528, total 1498482912." }, { "label": "RedHatAI Llama 3.2 1B Instruct FP8 config", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8/raw/fb49430ca7e61099fa1ff30f12fecb290a8ebb65/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, bfloat16, hidden size 2048, intermediate size 8192, 16 layers, 32 attention heads, 8 KV heads, 64 head dimension, 131072 max position embeddings, rope_theta 500000, Llama 3 rope scaling from original 8192 to factor 32, vocab size 128256, tie_word_embeddings true, and compressed-tensors FP8 static tensor activation and weight quantization with lm_head ignored." }, { "label": "RedHatAI Llama 3.2 1B Instruct FP8 recipe", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8/raw/fb49430ca7e61099fa1ff30f12fecb290a8ebb65/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier with targets [Linear], scheme FP8, observer mse, and ignore [lm_head]." }, { "label": "RedHatAI Llama 3.2 1B Instruct FP8 safetensors header", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8/resolve/fb49430ca7e61099fa1ff30f12fecb290a8ebb65/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "Range-reading the single safetensors header found 371 tensors totaling 2.023887296 GB: 0.973078528 GB F8_E4M3 tensors and 1.050808768 GB BF16 tensors. model.embed_tokens.weight is BF16 with shape [128256, 2048] and contributes 262668288 tensor elements / 0.525336576 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, including 112 BF16 input_scale tensors and 112 BF16 weight_scale tensors, plus model.norm.weight plus lm_head.weight total 1.49855072 GB swept traffic." }, { "label": "Meta Llama 3.2 1B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The base API metadata identifies the gated Meta repo as the public base model. Raw config access for the base repo is gated in this audit environment, so the quantized repo's public config is the direct geometry evidence for this profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from RedHatAI API metadata, served static FP8 config, recipe.yaml, safetensors header byte grouping, and Meta base API metadata." }, "notes": "This profile supersedes the scraped metadata estimate, which treated all stored weights as a flat 1 byte per parameter and missed BF16 resident/swept tensors plus static input/weight scale tensors stored by the compressed-tensors package." }, { "id": "redhatai--meta-llama-3-1-8b-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Meta-Llama-3.1-8B-FP8", "title": "RedHatAI Meta Llama 3.1 8B FP8", "summary": "Audited memory-side bounds profile for the RedHatAI compressed-tensors static FP8 package of Meta Llama 3.1 8B.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-8B", "relation": "quantized", "source": "Hugging Face model metadata, served quantized config, quantization recipe, and safetensors index/header range reads", "config_compatible": true, "notes": "The RedHatAI model card and API metadata record meta-llama/Meta-Llama-3.1-8B as the base model. The base Meta repo raw config is gated in this audit environment, but the public quantized config directly records the LlamaForCausalLM geometry used by this profile and its _name_or_path points at a Meta-Llama-3.1-8B snapshot." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261696, "swept_params_b": 7.50492512, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 9.081201536, "swept_weight_gb": 8.030528384, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "safetensors_index_total_size_and_range_read_stored_fp8_f16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings false, and the safetensors index stores separate model.embed_tokens.weight and lm_head.weight F16 tensors. The input embedding is resident-only for ordinary decode; the separate lm_head output projection remains in swept decode traffic. This static FP8 artifact also stores per-tensor input_scale and weight_scale tensors for quantized Linear modules, and those tiny F16 scale tensors are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 8 KV heads, 128 inferred head dimension from hidden size 4096 over 32 attention heads, and 131072 max position embeddings with Llama 3 RoPE scaling." }, "notes": "Dense LlamaForCausalLM profile using the served RedHatAI static FP8 config and exact stored safetensors bytes." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-static-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus F16 embeddings, lm_head, norms, input_scale tensors, and weight_scale tensors from safetensors headers. Static activation quantization, FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and compressed-tensors float-quantized FP8 weights with static tensor activation quantization, static tensor weight quantization, and lm_head ignored. kv_cache_scheme is null, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Meta Llama 3.1 8B FP8 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Meta-Llama-3.1-8B-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format" ], "notes": "At commit 881d74b74ad8ebe8775fb178b1fd0c935477f1bc, the API records a public llama3.1-licensed text-generation repo with base_model meta-llama/Llama-3.1-8B, compressed-tensors, region:us, and 238957 downloads. The API safetensors block reports tensor elements split across F16: 1050939840 and F8_E4M3: 6979321856, total 8030261696." }, { "label": "RedHatAI Meta Llama 3.1 8B FP8 model card", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "quantized_module_scope" ], "notes": "The model card describes this as a quantized base version of Meta-Llama-3.1-8B for vLLM, with FP8 weight and activation quantization. It states that only Linear operators inside transformer blocks are quantized, using symmetric per-tensor quantization and LLM Compressor calibration on 512 UltraChat sequences." }, { "label": "RedHatAI Meta Llama 3.1 8B FP8 config", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8/raw/881d74b74ad8ebe8775fb178b1fd0c935477f1bc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, float16, hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, 131072 max position embeddings, rope_theta 500000, Llama 3 RoPE scaling from original 8192 by factor 8, vocab size 128256, tie_word_embeddings false, and compressed-tensors FP8 static tensor activation and weight quantization with lm_head ignored and kv_cache_scheme null." }, { "label": "RedHatAI Meta Llama 3.1 8B FP8 recipe", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8/raw/881d74b74ad8ebe8775fb178b1fd0c935477f1bc/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier with targets [Linear], FP8 float weights, FP8 float input_activations, static tensor strategy, symmetric quantization, and ignore [lm_head]." }, { "label": "RedHatAI Meta Llama 3.1 8B FP8 safetensors index and headers", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8/raw/881d74b74ad8ebe8775fb178b1fd0c935477f1bc/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 9081201536 bytes across two safetensors shards. Range-reading both shard headers found 739 tensors totaling exactly 9.081201536 GB: 6.979321856 GB F8_E4M3 tensors and 2.101879680 GB F16 tensors. model.embed_tokens.weight is F16 with shape [128256, 4096] and contributes 525336576 tensor elements / 1.050673152 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, model.norm.weight, lm_head.weight, 224 input_scale tensors, and 224 weight_scale tensors total 8.030528384 GB swept traffic. Linked-object HEAD checks resolved both shards to 9081286496 total bytes, leaving 84960 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 config comparison", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/config.json", "source_type": "manual_review", "supports": [ "model_family", "kv_adapter", "quantized_module_scope" ], "notes": "Manual comparison against the already audited RedHatAI static FP8 Llama 3.1 8B Instruct config found matching architecture, model type, hidden size, intermediate size, layer count, attention heads, KV heads, max-position embeddings, RoPE scaling, RoPE theta, vocabulary size, tied-embedding setting, cache setting, quantization groups, ignored lm_head, and null KV cache scheme. This base config records torch_dtype float16 and quantization_status compressed; the Instruct sibling records torch_dtype bfloat16 and quantization_status frozen." }, { "label": "Meta Llama 3.1 8B API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-8B", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The existing base profile records the Meta repo as gated and not directly config-readable in this audit environment. This RedHatAI quantized repo therefore uses its own public config and tensor headers as the direct geometry evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current RedHatAI API metadata, model card, served static FP8 config, recipe.yaml, safetensors index, linked-object HEAD checks, direct shard header range reads, and comparison against the existing RedHatAI Llama 3.1 8B Instruct FP8 profile." }, "notes": "This profile supersedes the scraped metadata estimate, which treated all stored weights as a flat 1 byte per parameter and missed the F16 resident/swept tensors plus static input/weight scale tensors stored by the compressed-tensors package." }, { "id": "redhatai--meta-llama-3-1-8b-instruct-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", "title": "RedHatAI Meta Llama 3.1 8B Instruct FP8", "summary": "Audited memory-side bounds profile for the RedHatAI compressed-tensors static FP8 package of Meta Llama 3.1 8B Instruct.", "model_family": "llama3.1-dense", "base_model_proof": { "base_model": "meta-llama/Meta-Llama-3.1-8B-Instruct", "relation": "quantized", "source": "Hugging Face model metadata, served quantized config, quantization recipe, and safetensors index/header range reads", "config_compatible": true, "notes": "The RedHatAI model card records meta-llama/Meta-Llama-3.1-8B-Instruct as the base model, while the API tags also include the common meta-llama/Llama-3.1-8B-Instruct spelling. The base repo raw config is gated in this audit environment, but the public quantized config directly records the LlamaForCausalLM geometry used by this profile." }, "architecture": { "canonical_architecture_id": "llama-3-1-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261696, "swept_params_b": 7.50492512, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 9.081201536, "swept_weight_gb": 8.030528384, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "safetensors_index_total_size_and_range_read_stored_fp8_bf16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but not swept once decode has token embeddings", "notes": "The config records tie_word_embeddings false, and the safetensors index stores separate model.embed_tokens.weight and lm_head.weight BF16 tensors. The input embedding is resident-only for ordinary decode; the separate lm_head output projection remains in swept decode traffic. This static FP8 artifact also stores per-tensor input_scale and weight_scale tensors for quantized Linear modules, and those tiny scale tensors are included in swept layer traffic." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records 32 layers, 8 KV heads, 128 head dimension, and 131072 max position embeddings with Llama 3 RoPE scaling." }, "notes": "Dense LlamaForCausalLM profile using the served RedHatAI static FP8 config and exact stored safetensors bytes." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-static-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 weights plus BF16 embeddings, lm_head, norms, input_scale tensors, and weight_scale tensors from safetensors headers. Static activation quantization, FP8 dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype bfloat16 and compressed-tensors float-quantized FP8 weights with static tensor activation quantization, static tensor weight quantization, and lm_head ignored. kv_cache_scheme is null, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "weight_format" ], "notes": "At commit 12fd6884d2585dd4d020373e7f39f74507b31866, the API records a public llama3.1-licensed text-generation repo with base_model meta-llama/Meta-Llama-3.1-8B-Instruct, compressed-tensors, region:us, and 534098 downloads. The API safetensors block reports tensor elements split across BF16: 1050939840 and F8_E4M3: 6979321856, total 8030261696." }, { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 model card", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", "source_type": "model_card", "supports": [ "base_model_proof", "weight_format", "serving", "quantized_module_scope" ], "notes": "The model card describes this as a quantized version of Meta-Llama-3.1-8B-Instruct for vLLM, with FP8 weight and activation quantization. It states that only Linear operators inside transformer blocks are quantized, using symmetric per-tensor quantization and LLM Compressor calibration on 512 UltraChat sequences." }, { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 config", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, bfloat16, hidden size 4096, intermediate size 14336, 32 layers, 32 attention heads, 8 KV heads, 131072 max position embeddings, rope_theta 500000, Llama 3 RoPE scaling from original 8192 by factor 8, vocab size 128256, tie_word_embeddings false, and compressed-tensors FP8 static tensor activation and weight quantization with lm_head ignored and kv_cache_scheme null." }, { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 recipe", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier with targets [Linear], FP8 float weights, FP8 float input_activations, static tensor strategy, symmetric quantization, and ignore [lm_head]." }, { "label": "RedHatAI Meta Llama 3.1 8B Instruct FP8 safetensors index and headers", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "The index records total_size 9081201536 bytes across two safetensors shards. Range-reading both shard headers found 739 tensors totaling exactly 9.081201536 GB: 6.979321856 GB F8_E4M3 tensors and 2.101879680 GB BF16 tensors. model.embed_tokens.weight is BF16 with shape [128256, 4096] and contributes 525336576 tensor elements / 1.050673152 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, model.norm.weight, lm_head.weight, 224 input_scale tensors, and 224 weight_scale tensors total 8.030528384 GB swept traffic. Linked-object HEAD checks resolved both shards to 9081287016 total bytes, leaving 85480 bytes of safetensors header/container overhead outside tensor payloads." }, { "label": "Meta Llama 3.1 8B Instruct API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-8B-Instruct", "source_type": "model_card", "supports": [ "base_model_proof" ], "notes": "The existing base profile records the Meta repo as gated and not directly config-readable in this audit environment. This RedHatAI quantized repo therefore uses its own public config and tensor headers as the direct geometry evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current RedHatAI API metadata, model card, served static FP8 config, recipe.yaml, safetensors index, linked-object HEAD checks, direct shard header range reads, and the existing gated Meta base profile." }, "notes": "This profile supersedes the scraped metadata estimate, which treated all stored weights as a flat 1 byte per parameter and missed the BF16 resident/swept tensors plus static input/weight scale tensors stored by the compressed-tensors package." }, { "id": "redhatai--qwen2-5-1-5b-quantized-w8a8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen2.5-1.5B-quantized.w8a8", "title": "RedHatAI Qwen2.5 1.5B W8A8", "summary": "Audited memory-side bounds profile for the RedHatAI compressed-tensors INT8 weight/activation package of Qwen2.5 1.5B.", "model_family": "qwen2.5-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-1.5B", "relation": "quantized", "source": "Hugging Face model card, served quantized config, recipe, and source-snapshot config comparison", "config_compatible": true, "notes": "The RedHatAI repo records Qwen/Qwen2.5-1.5B as its base model. The quantized config's _name_or_path points to Qwen snapshot cf341214a7c12d36db6c7fbbf8b113c7ed61502f; manual comparison found matching tensor geometry, dtype label, max context, vocab size, attention geometry, and tied-embedding setting. The quantized config adds compressed-tensors metadata and changes disabled sliding_window from 32768 to null, which does not affect KV because use_sliding_window is false." }, "architecture": { "canonical_architecture_id": "qwen2-5-1-5b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.77773312, "swept_params_b": 1.544359424, "auxiliary_resident_params_b": 0.233373696, "resident_weight_gb": 2.245270528, "swept_weight_gb": 1.778523136, "auxiliary_resident_weight_gb": 0.466747392, "resident_parameter_scope": "safetensors_header_stored_int8_bf16_tensor_elements", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, quantization scales/biases, and lm_head.weight output projection", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "The config records tie_word_embeddings true, but the safetensors file stores both model.embed_tokens.weight and lm_head.weight separately. This profile charges the stored resident artifact and keeps lm_head.weight in swept output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 28 layers, 2 KV heads, hidden size 1536, 12 attention heads, and use_sliding_window false. Head dimension is 1536 / 12 = 128. The quantization_config has kv_cache_scheme null, so KV cache is charged at BF16." }, "notes": "Dense Qwen2ForCausalLM profile using the served RedHatAI W8A8 config directly. The current upstream Qwen/Qwen2.5-1.5B config has since moved to 131072 max positions, but this quantized artifact preserves the earlier 32768-position source snapshot." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w8a8-memory-bound", "dequantization_notes": "The memory-side bound charges stored I8 linear weights plus BF16 embeddings, lm_head, norms, biases, and per-channel weight_scale tensors from the safetensors header. Dynamic INT8 activation quantization, dequantization compute, and activation-scale traffic are outside Bounds Engine v1.", "notes": "The model card and recipe describe W8A8 quantization: static symmetric per-channel INT8 weights and dynamic symmetric per-token INT8 activations for Linear targets. The config keeps KV cache unquantized with kv_cache_scheme null." }, "evidence": [ { "label": "RedHatAI Qwen2.5 1.5B W8A8 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-1.5B-quantized.w8a8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "total_params_b", "weight_format" ], "notes": "At commit 45d7501a9947504178b3d1bfee36272bd1582e66, the API records a public Apache-2.0 text-generation repo with base_model Qwen/Qwen2.5-1.5B, compressed-tensors, 8-bit, llmcompressor, neuralmagic, and region:us tags; 1009174 downloads; and safetensors parameters BF16 467537408, I8 1310195712, total 1777733120. The model card says only Linear operators inside transformer blocks are quantized, with INT8 weights and INT8 activations." }, { "label": "RedHatAI Qwen2.5 1.5B W8A8 config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/raw/45d7501a9947504178b3d1bfee36272bd1582e66/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "tie_word_embeddings" ], "notes": "The config records Qwen2ForCausalLM, qwen2, bfloat16 runtime dtype, hidden size 1536, intermediate size 8960, 28 layers, 12 attention heads, 2 KV heads, 32768 max position embeddings, max_window_layers 28, use_sliding_window false, vocab size 151936, rope_theta 1000000, tie_word_embeddings true, and compressed-tensors int-quantized W8A8 Linear targets with lm_head ignored and kv_cache_scheme null." }, { "label": "Qwen2.5 1.5B source snapshot config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/cf341214a7c12d36db6c7fbbf8b113c7ed61502f/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching architecture, dtype label, hidden size, intermediate size, layer count, attention head count, KV head count, max_position_embeddings, max_window_layers, use_sliding_window, tied embeddings, vocab size, rope_theta, rms_norm_eps, and use_mrope. The only audited difference is disabled sliding_window 32768 in the source snapshot versus null in the RedHatAI quantized config." }, { "label": "RedHatAI Qwen2.5 1.5B W8A8 quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/raw/45d7501a9947504178b3d1bfee36272bd1582e66/recipe.yaml", "source_type": "manual_review", "supports": [ "weight_format", "serving" ], "notes": "The recipe records a GPTQModifier with scheme W8A8, Linear targets, sequential_update true, dampening_frac 0.01, observer mse, and ignore [lm_head]." }, { "label": "RedHatAI Qwen2.5 1.5B W8A8 safetensors header", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/resolve/45d7501a9947504178b3d1bfee36272bd1582e66/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "Range-reading the single safetensors header found a 60952-byte header with 535 tensors totaling 2245270528 payload bytes: I8 1310195712 bytes and BF16 935074816 bytes. The linked object size is 2245331488 bytes. model.embed_tokens.weight is BF16 with shape [151936, 1536] and contributes 233373696 tensor elements / 0.466747392 GB resident-only for ordinary decode. lm_head.weight is stored separately with the same shape and remains in swept decode traffic. Layer tensors, model.norm.weight, quantization weight_scale/bias tensors, and lm_head.weight total 1.544359424B tensor elements / 1.778523136 GB swept traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from RedHatAI API metadata, model card, served compressed-tensors config, quantization recipe, source-snapshot config comparison, direct safetensors header byte grouping, and the existing local Qwen2.5 1.5B profile." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the artifact as a flat 1-byte dense model and missed the BF16 embeddings, lm_head, norms, biases, and quantization-scale tensors stored by the compressed-tensors package." }, { "id": "redhatai--qwen2-5-vl-3b-instruct-quantized-w4a16", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16", "title": "RedHatAI Qwen2.5 VL 3B Instruct W4A16", "summary": "Audited memory-side text-decode bounds profile for RedHatAI's compressed-tensors W4A16 package of Qwen2.5-VL 3B Instruct.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-3B-Instruct", "relation": "quantized", "source": "Hugging Face API metadata, model card, served compressed-tensors config, recipe.yaml, BF16 base config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The repo metadata and model card identify Qwen/Qwen2.5-VL-3B-Instruct as the quantized base model. Manual comparison found matching checked text geometry and context fields; the RedHatAI artifact adds compressed-tensors W4A16 quantization metadata and truncates vision_config metadata while preserving the resident BF16 visual tensors in the safetensors header." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-3b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.08746444, "swept_params_b": 3.107615224, "auxiliary_resident_params_b": 0.979849216, "resident_weight_gb": 4.01313376, "swept_weight_gb": 2.053435328, "auxiliary_resident_weight_gb": 1.959698432, "resident_parameter_scope": "logical Qwen2.5-VL 3B parameter count represented by the compressed-tensors safetensors package", "swept_parameter_scope": "ordinary text decode charges model.layers.*, model.norm.weight, and the separate BF16 lm_head.weight stored by this artifact", "auxiliary_scope": "visual tensors plus model.embed_tokens.weight are resident for the multimodal package but are not swept as full matrices for each ordinary generated text token", "notes": "Bounds use exact stored bytes from the single safetensors header because the package mixes packed I32 weight_packed tensors, BF16 weight_scale/bias tensors, BF16 embeddings, BF16 lm_head, BF16 visual tensors, and I64 weight_shape tensors. The config records tie_word_embeddings true, but this compressed-tensors artifact stores a separate lm_head.weight and recipe.yaml explicitly ignores lm_head, so ordinary text decode excludes only the input embedding table while keeping lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.9818149659547863, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w4a16-qwen2.5-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: packed I32 weights, BF16 scales/biases, BF16 ignored modules, and I64 shape tensors from the safetensors header. Dequantization, activation traffic, vision encoder throughput, and compute overhead are outside Bounds Engine v1.", "notes": "The config records bfloat16 and compressed-tensors pack-quantized 4-bit integer weights with group_size 128, symmetric group quantization, weight act-order, and kv_cache_scheme null. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Qwen2.5 VL 3B W4A16 API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit a6bca3776bd02c77a80e66326c4441a651fdd2e7, the API records a public non-gated Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen2.5-VL-3B-Instruct, vLLM, vision, W4A16, compressed-tensors, endpoints_compatible, region:us, and 115971 downloads. The API safetensors block reports I64: 504, I32: 2774532096 logical packed int4 parameters, BF16: 1312931840, and total: 4087464440." }, { "label": "RedHatAI Qwen2.5 VL 3B W4A16 model card", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "runtime_format" ], "notes": "The card identifies Qwen/Qwen2.5-VL-3B-Instruct as the architecture and base model, describes INT4 weight quantization with FP16/BF16 activation storage for vLLM, and shows llm-compressor creation with GPTQModifier, 512 calibration samples, max calibration sequence length 2048, sequential Qwen2_5_VLDecoderLayer targets, and ignored lm_head plus visual modules." }, { "label": "RedHatAI Qwen2.5 VL 3B W4A16 served config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/raw/a6bca3776bd02c77a80e66326c4441a651fdd2e7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, bfloat16, compressed-tensors pack-quantized 4-bit integer weights, group_size 128, symmetric group strategy, weight act-order, kv_cache_scheme null, 36 text layers, hidden size 2048, intermediate size 11008, 16 attention heads, 2 KV heads, 128000 max position embeddings, tie_word_embeddings true, use_sliding_window false, mRoPE, vocab size 151936, and a resident visual tower summarized by hidden_size 1280." }, { "label": "RedHatAI Qwen2.5 VL 3B W4A16 recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/raw/a6bca3776bd02c77a80e66326c4441a651fdd2e7/recipe.yaml", "source_type": "config", "supports": [ "quantization", "swept_parameter_scope", "auxiliary_resident_scope" ], "notes": "recipe.yaml records GPTQModifier with Qwen2_5_VLDecoderLayer sequential targets, dampening_frac 0.01, 4-bit int symmetric group quantization, group_size 128, strategy group, actorder weight, and ignored lm_head plus visual modules." }, { "label": "Qwen2.5 VL 3B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "auxiliary_resident_scope" ], "notes": "Manual comparison found matching architecture, model_type, dtype, tie_word_embeddings, text layer count, hidden size, intermediate size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, rope_theta, and mRoPE section between the RedHatAI config and the BF16 Instruct config. The RedHatAI config adds quantization_config and truncates detailed vision_config fields, but the base config records visual depth 32, hidden_size 1280, intermediate_size 3420, 16 heads, patch_size 14, spatial_merge_size 2, and full-attention blocks 7/15/23/31." }, { "label": "RedHatAI Qwen2.5 VL 3B W4A16 safetensors header", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/resolve/a6bca3776bd02c77a80e66326c4441a651fdd2e7/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "Direct HTTP range reads found a 148768-byte safetensors header, 1329 tensors, and a linked file size of 4.013282536 GB. Tensor payload bytes total 4.013133760 GB, leaving 0.000148776 GB of header/container overhead outside tensor payloads. Stored bytes split into BF16 2.625863680 GB, I32 1.387266048 GB, and I64 0.000004032 GB. Suffix groups are weight_packed 1.387266048 GB, weight_scale 0.043352064 GB, BF16 weight 2.581465600 GB, BF16 bias 0.001046016 GB, and weight_shape 0.000004032 GB. model.layers.* tensors total 1.431101376 GB, model.norm.weight 0.000004096 GB, lm_head.weight 0.622329856 GB, model.embed_tokens.weight 0.622329856 GB, and visual.* tensors 1.337368576 GB. Ordinary text swept traffic is model.layers plus model.norm plus lm_head, totaling 2.053435328 GB; resident-only visual plus input embedding tensors total 1.959698432 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, recipe.yaml, current BF16 base config comparison, and direct range-read safetensors header byte grouping." }, "notes": "This profile supersedes the generated ideal 4-bit dense estimate, which undercounted compressed-tensors side tensors and missed the separate BF16 lm_head plus resident-only BF16 input embedding and visual tower split." }, { "id": "redhatai--qwen2-5-vl-3b-instruct-quantized-w8a8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8", "title": "RedHatAI Qwen2.5 VL 3B Instruct W8A8", "summary": "Audited memory-side text-decode bounds profile for RedHatAI's compressed-tensors W8A8 package of Qwen2.5-VL 3B Instruct.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-3B-Instruct", "relation": "quantized", "source": "Hugging Face API metadata, model card, served compressed-tensors config, recipe.yaml, BF16 base config comparison, and direct safetensors header metadata", "config_compatible": true, "notes": "The repo metadata and model card identify Qwen/Qwen2.5-VL-3B-Instruct as the quantized base model. Manual comparison found matching checked text geometry, full vision_config metadata, and context fields; the RedHatAI artifact adds compressed-tensors W8A8 quantization metadata while preserving BF16 lm_head, BF16 input embedding, and BF16 resident visual tensors." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-3b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.066820096, "swept_params_b": 3.08697088, "auxiliary_resident_params_b": 0.979849216, "resident_weight_gb": 5.359108096, "swept_weight_gb": 3.399409664, "auxiliary_resident_weight_gb": 1.959698432, "resident_parameter_scope": "logical Qwen2.5-VL 3B parameter count represented by the compressed-tensors safetensors package", "swept_parameter_scope": "ordinary text decode charges model.layers.*, model.norm.weight, and the separate BF16 lm_head.weight stored by this artifact", "auxiliary_scope": "visual tensors plus model.embed_tokens.weight are resident for the multimodal package but are not swept as full matrices for each ordinary generated text token", "notes": "Bounds use exact stored bytes from both safetensors shard headers because the package mixes I8 quantized Linear weights, BF16 weight_scale tensors, BF16 biases, BF16 ignored modules, BF16 embeddings, BF16 lm_head, and BF16 visual tensors. The config records tie_word_embeddings true, but this compressed-tensors artifact stores a separate lm_head.weight and recipe.yaml explicitly ignores lm_head, so ordinary text decode excludes only the input embedding table while keeping lm_head.weight in swept traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records use_sliding_window false. The 128 head dimension is derived from hidden_size 2048 divided by 16 attention heads." }, "notes": "Qwen2_5_VLForConditionalGeneration is multimodal. This profile models ordinary text decode after any image/video prefill, not vision encoder throughput." }, "serving": { "weight_format": "int8", "weight_bytes_per_param": 1, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-w8a8-qwen2.5-vl-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored compressed-tensors bytes: I8 Linear weights, BF16 scales/biases, BF16 ignored modules, BF16 embeddings, BF16 lm_head, and BF16 visual tensors from the safetensors headers. Dynamic INT8 activation quantization, dequantization, activation traffic, vision encoder throughput, and compute overhead are outside Bounds Engine v1.", "notes": "The model card and recipe describe W8A8 quantization: static symmetric per-channel INT8 weights and dynamic symmetric per-token INT8 activations for Linear targets. The config keeps KV cache unquantized with kv_cache_scheme null, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "RedHatAI Qwen2.5 VL 3B W8A8 API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit fcd4bf97086d0a782ead353e4ea700db0628b568, the API records a public non-gated Apache-2.0 image-text-to-text repo with base_model Qwen/Qwen2.5-VL-3B-Instruct, vLLM, vision, W8A8, compressed-tensors, endpoints_compatible, region:us, and 121846 downloads. The API safetensors block reports I8: 2774532096 logical INT8 parameters, BF16: 1292288000, and total: 4066820096." }, { "label": "RedHatAI Qwen2.5 VL 3B W8A8 model card", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "runtime_format" ], "notes": "The card identifies Qwen/Qwen2.5-VL-3B-Instruct as the architecture and base model, describes INT8 weight quantization plus INT8 activation quantization for vLLM, and shows llm-compressor creation with GPTQModifier, 512 calibration samples, max calibration sequence length 2048, sequential Qwen2_5_VLDecoderLayer targets, and ignored lm_head plus visual modules." }, { "label": "RedHatAI Qwen2.5 VL 3B W8A8 served config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/raw/fcd4bf97086d0a782ead353e4ea700db0628b568/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "weight_format", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen2_5_VLForConditionalGeneration, bfloat16, compressed-tensors int-quantized W8A8 Linear weights/activations, static per-channel INT8 weights, dynamic per-token INT8 activations, kv_cache_scheme null, 36 text layers, hidden size 2048, intermediate size 11008, 16 attention heads, 2 KV heads, 128000 max position embeddings, tie_word_embeddings true, use_sliding_window false, mRoPE, vocab size 151936, and a resident 32-layer visual tower." }, { "label": "RedHatAI Qwen2.5 VL 3B W8A8 recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/raw/fcd4bf97086d0a782ead353e4ea700db0628b568/recipe.yaml", "source_type": "config", "supports": [ "quantization", "swept_parameter_scope", "auxiliary_resident_scope" ], "notes": "recipe.yaml records GPTQModifier with Qwen2_5_VLDecoderLayer sequential targets, W8A8 scheme, Linear targets, and ignored lm_head plus visual modules." }, { "label": "Qwen2.5 VL 3B Instruct base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "auxiliary_resident_scope" ], "notes": "Manual comparison found matching architecture, model_type, dtype, tie_word_embeddings, text layer count, hidden size, intermediate size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, rope_theta, mRoPE section, and vision_config between the RedHatAI config and the BF16 Instruct config. The base config records visual depth 32, hidden_size 1280, intermediate_size 3420, 16 heads, patch_size 14, spatial_merge_size 2, and full-attention blocks 7/15/23/31." }, { "label": "RedHatAI Qwen2.5 VL 3B W8A8 safetensors headers", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/resolve/fcd4bf97086d0a782ead353e4ea700db0628b568/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "weight_format", "embedding_layout" ], "notes": "Direct HTTP range reads found two safetensors shards, 1077 tensors, and linked shard sizes of 4.736898848 GB and 0.622329984 GB. Tensor payload bytes total 5.359108096 GB, leaving 0.000120736 GB of header/container overhead outside tensor payloads. Stored bytes split into I8 2.774532096 GB and BF16 2.584576000 GB. Suffix groups are weight 5.355997696 GB, weight_scale 0.002064384 GB, and bias 0.001046016 GB. model.layers.* tensors total 2.777075712 GB, model.norm.weight 0.000004096 GB, lm_head.weight 0.622329856 GB, model.embed_tokens.weight 0.622329856 GB, and visual.* tensors 1.337368576 GB. Ordinary text swept traffic is model.layers plus model.norm plus lm_head, totaling 3.399409664 GB; resident-only visual plus input embedding tensors total 1.959698432 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors config, recipe.yaml, current BF16 base config comparison, and direct range-read safetensors header byte grouping across both shards." }, "notes": "This profile supersedes the generated ideal 8-bit dense estimate, which undercounted compressed-tensors side tensors and missed the separate BF16 lm_head plus resident-only BF16 input embedding and visual tower split." }, { "id": "redhatai--qwen3-5-122b-a10b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen3.5-122B-A10B-NVFP4", "title": "RedHatAI Qwen3.5 122B A10B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI compressed-tensors NVFP4 Qwen3.5 122B A10B serving artifact.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, compression recipe, audited BF16 base config comparison, and direct safetensors header grouping", "config_compatible": true, "notes": "The RedHatAI artifact records Qwen/Qwen3.5-122B-A10B as its base model and preserves the audited Qwen3.5 text geometry: Qwen3_5MoeForConditionalGeneration, 48 text layers, 12 full-attention layers, 36 DeltaNet linear-attention layers, 2 KV heads, 256 full-attention head dimension, 256 experts, 8 routed experts per token, one shared expert, 262144 max positions, and the same DeltaNet state geometry. The quantized repo adds compressed-tensors NVFP4 metadata. The README text says Qwen3NextForCausalLM, but the served config, safetensors names, and memory-relevant geometry are Qwen3.5." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 76.419752928, "main_resident_weight_gb": 73.99149312, "auxiliary_resident_weight_gb": 2.428259808, "fixed_weight_gb": 8.7613824, "routed_expert_weight_gb": 0.25480512, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_compressed_tensors_nvfp4_u8_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding and vision tower tensors; no MTP tensors were present in this artifact", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024. The compression recipe ignores shared_expert_gate but not the shared expert Linear modules, so shared expert traffic is included in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and tiny F32 scale tensors. Routed experts are byte-uniform across all 256 expert indexes; routed_expert_weight_gb is the grouped ordinary-language routed tensor byte count divided by 256." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with full_attention_interval 4, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The RedHatAI compressed-tensors config has kv_cache_scheme null, so this profile does not inherit the separate txn545 ModelOpt FP8-KV assumption." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-llm-compressor-nvfp4-bf16-kv-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors NVFP4/FP8/BF16/F32 safetensors bytes and BF16 KV bytes. NVFP4 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, and state writes are outside this memory-side bound.", "notes": "The served config records compressed-tensors nvfp4-pack-quantized weights and activations with kv_cache_scheme null. weight_bytes_per_param records the nominal NVFP4 weight payload; the audited adapter uses exact stored tensor bytes for resident and per-token weight traffic." }, "evidence": [ { "label": "RedHatAI Qwen3.5 122B A10B NVFP4 API metadata and model card", "url": "https://huggingface.co/api/models/RedHatAI/Qwen3.5-122B-A10B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "weight_format", "serving", "total_params_b" ], "notes": "At commit 49d19c108259a21450c40b8af38828b0a97390d8, the API records a public non-gated compressed-tensors artifact derived from Qwen/Qwen3.5-122B-A10B, with current downloads 246347, qwen/nvfp4/vllm/compressed-tensors tags, region:us, usedStorage 76477963570 bytes, and safetensors logical parameters BF16 5201997552, F8_E4M3 7335051264, U8 58680410112, F32 74112, total 71217533040. The target repo does not currently declare a license or pipeline_tag in its API/card metadata." }, { "label": "RedHatAI Qwen3.5 122B A10B NVFP4 config", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/raw/49d19c108259a21450c40b8af38828b0a97390d8/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "kv_store_format", "kv_read_format", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 48 text layers, full_attention_interval 4, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, compressed-tensors nvfp4-pack-quantized weights and activations, and kv_cache_scheme null." }, { "label": "RedHatAI Qwen3.5 122B A10B NVFP4 compression recipe", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/raw/49d19c108259a21450c40b8af38828b0a97390d8/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "ignored_quantized_modules", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "The LLM Compressor recipe targets Linear modules with NVFP4, while ignoring lm_head, visual modules, mlp.gate, embed_tokens, shared_expert_gate, and linear_attn modules. The resulting stored bytes leave always-on language traffic materially larger than a flat 0.5-byte parameter estimate." }, { "label": "Qwen3.5 122B A10B base config and API comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "license" ], "notes": "Manual comparison found matching audited top-level, text, vision, MoE, attention, and DeltaNet state geometry fields between the RedHatAI NVFP4 config and the BF16 base config. The base API records Apache-2.0 metadata and safetensors parameters BF16 125086490096 plus F32 6912." }, { "label": "RedHatAI Qwen3.5 122B A10B NVFP4 safetensors headers", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/resolve/49d19c108259a21450c40b8af38828b0a97390d8/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across both indexed shards. The index records total_size 76419752928 bytes; direct headers found 149100 tensors totaling 76.419752928 GB, split into 58.680410112 GB U8, 7.335051264 GB F8_E4M3, 10.403995104 GB BF16, and 0.000296448 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 73.991493120 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 2.428259808 GB. No MTP tensors were present under the audited prefixes. Ordinary-language routed expert tensors sum to 65.230110720 GB and divide exactly into 256 uniform expert indexes of 0.254805120 GB. Fixed ordinary text traffic sums to 8.761382400 GB. All index weight_map entries were found in the shard headers." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, served config, compression recipe, BF16 base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It separates resident visual/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. It deliberately differs from txn545/Qwen3.5-122B-A10B-NVFP4 because this RedHatAI compressed-tensors artifact does not declare FP8 KV." }, { "id": "redhatai--qwen3-5-9b-fp8-dynamic", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen3.5-9B-FP8-dynamic", "title": "RedHatAI Qwen3.5 9B FP8 Dynamic", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI FP8-dynamic compressed-tensors package of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face API metadata, pinned compressed-tensors config and recipe, base config comparison, and safetensors index/header range reads", "config_compatible": true, "notes": "The API metadata identifies Qwen/Qwen3.5-9B as the quantized base model. Manual comparison found matching checked top-level and text architecture fields: Qwen3_5ForConditionalGeneration, 32 text layers, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, one MTP layer, and 262144 max position embeddings. The package adds compressed-tensors FP8 dynamic quantization metadata while preserving the base hybrid attention geometry." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.653104368, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.716419824, "resident_weight_gb": 14.006672864, "swept_weight_gb": 10.573833216, "auxiliary_resident_weight_gb": 3.432839648, "resident_parameter_scope": "base logical Qwen3.5 9B parameters with direct compressed-tensors FP8/BF16 safetensors stored-byte totals", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from the two safetensors shard headers because this compressed-tensors package mixes F8_E4M3 self-attention and MLP matrix weights with BF16 scales, BF16 linear-attention tensors, BF16 embeddings, BF16 lm_head, BF16 vision, and BF16 MTP tensors. Logical parameter counts follow the audited Qwen3.5 BF16 profile so model identity remains the 9.653104368B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 32 layers with every fourth layer marked full_attention, giving 8 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The compressed-tensors artifact preserves the base Qwen3.5 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, dequantization, dynamic activation quantization, and runtime scheduler behavior remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The config records kv_cache_scheme null, so KV cache is charged at BF16." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.4510019088193082, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-fp8-dynamic-qwen3.5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors safetensors bytes: F8_E4M3 self-attention and MLP matrix weights plus BF16 side tensors and unquantized modules. Dynamic per-token activation quantization, dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors float quantization with static per-channel FP8 weights, dynamic per-token FP8 input activations, and kv_cache_scheme null. The recipe ignores lm_head, embed_tokens, visual modules, and linear-attention modules; the config ignore list also excludes top-level MTP modules. The resulting quantized storage scope is ordinary language self-attention and MLP Linear modules." }, "evidence": [ { "label": "RedHatAI Qwen3.5 9B FP8 Dynamic API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen3.5-9B-FP8-dynamic", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 790f0576d2d77dd5322aa0603a470bd9e3a3d1f6, the live API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.5-9B, with fp8, vllm, llm-compressor, compressed-tensors, endpoints_compatible, and region:us tags. Current downloads are 178278. The API safetensors block reports BF16: 4109245680, F8_E4M3: 5301600256, and total: 9410845936 storage-accounting elements." }, { "label": "RedHatAI Qwen3.5 9B FP8 Dynamic config", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors float quantization, static per-channel FP8 weights, dynamic per-token FP8 activations, text_config model_type qwen3_5_text, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, one MTP layer, and kv_cache_scheme null." }, { "label": "RedHatAI Qwen3.5 9B FP8 Dynamic quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/recipe.yaml", "source_type": "config", "supports": [ "serving", "quantization", "quantized_module_scope" ], "notes": "The recipe records QuantizationModifier targets [Linear], scheme FP8_DYNAMIC, and ignores lm_head, embed_tokens, visual modules, and linear_attn modules. The pinned config adds re:^mtp.* to the ignored module list. Header grouping confirms that language self-attention and MLP Linear weights are F8_E4M3 while language linear-attention, embeddings, lm_head, visual, and MTP tensors remain BF16." }, { "label": "Qwen3.5 9B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof", "linear_attention_state" ], "notes": "Manual comparison found matching top-level identity fields and text geometry fields between the base BF16 repo and this compressed-tensors FP8 dynamic artifact. Both configs record the same hybrid full-attention/linear-attention layer pattern, resident vision tower, and one MTP layer." }, { "label": "RedHatAI Qwen3.5 9B FP8 Dynamic safetensors headers", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 14006785856 bytes across model.safetensors and model_mtp.safetensors. Direct range reads of both shard headers found 903 tensors totaling exactly 14.006672864 GB: BF16 8.705072608 GB and F8_E4M3 5.301600256 GB. Linked-object HEAD checks resolved both shards to 14.006785856 GB, leaving 112992 bytes of safetensors header/container overhead outside tensor payloads. The swept ordinary text subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 10.573833216 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 3.432839648 GB. Header buckets are language MLP tensors 4.833673216 GB, language linear-attention tensors 3.235390464 GB, lm_head 2.034237440 GB, input embedding 2.034237440 GB, visual 0.912020960 GB, MTP 0.486581248 GB, language self-attention tensors 0.469999616 GB, language layer/norm tensors 0.000524288 GB, and final language norm 0.000008192 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned compressed-tensors config and recipe, current base config comparison, direct range-read safetensors headers, linked-object HEAD checks, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 1-byte dense weights and all-layer full-context KV. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept text-decode weights and includes the fixed DeltaNet runtime-state charge." }, { "id": "redhatai--qwen3-6-35b-a3b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "RedHatAI/Qwen3.6-35B-A3B-NVFP4", "title": "RedHatAI Qwen3.6 35B A3B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the RedHatAI NVFP4 Qwen3.6 35B A3B serving artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, and safetensors header review", "config_compatible": true, "notes": "The RedHatAI artifact records Qwen/Qwen3.6-35B-A3B as its base model and preserves the Qwen3.6 text geometry: 40 text layers, 256 routed experts, 8 routed experts per token, shared_expert_intermediate_size 512, and every fourth layer using full attention." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 25.025991648, "main_resident_weight_gb": 21.426448896, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 3.306809856, "routed_expert_weight_gb": 0.07077984, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model_visual.safetensors tensors, model_mtp.safetensors tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The RedHatAI quantization recipe leaves linear_attn modules, mlp.gate, shared_expert_gate, lm_head, embed_tokens, visual tensors, and MTP tensors out of the NVFP4 Linear target set; this makes fixed ordinary text traffic materially larger than the NVIDIA NVFP4 profile. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The RedHatAI NVFP4 artifact preserves the base Qwen3.6 text architecture, so quantizing weights and activations does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head, with resident-only multimodal and MTP tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-llm-compressor-nvfp4-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored NVFP4 packed weights, FP8 scale tensors, BF16 unquantized tensors, F32 scalar scales, and BF16 KV bytes. Activation quantization, dequantization, compute overhead, and write traffic are outside Bounds Engine v1.", "notes": "The config records compressed-tensors nvfp4-pack-quantized weights and activations with kv_cache_scheme null. The model card serving command does not request FP8 KV cache, so this profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "RedHatAI Qwen3.6 35B A3B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen3.6-35B-A3B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "weight_format", "total_params_b", "serving" ], "notes": "At commit e7759a52191808ad69737acd5d93f6f78c1af4f5, the API reports an Apache-2.0 repo with base_model Qwen/Qwen3.6-35B-A3B, safetensors dtype counts F32 61760, BF16 2049897088, F8_E4M3 2038169600, and U8 16305356800, plus qwen3_5_moe, nvfp4, vllm, and compressed-tensors tags. The card states this is a preliminary NVFP4 quantization with weights and activations quantized through vllm-project/llm-compressor, and its examples serve the language model with vLLM." }, { "label": "RedHatAI Qwen3.6 35B A3B NVFP4 config", "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4/raw/e7759a52191808ad69737acd5d93f6f78c1af4f5/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "quantization_ignore_scope" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration, a qwen3_5_moe_text text config, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, and kv_cache_scheme null. The quantization_config targets Linear but ignores visual modules, all linear_attn projections, mlp.gate, shared_expert_gate, lm_head, embed_tokens, and mtp tensors." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison found the audited Qwen3.6 text geometry preserved between the base repo and the RedHatAI NVFP4 artifact. The RedHatAI artifact adds compressed-tensors quantization while preserving the base text architecture." }, { "label": "RedHatAI Qwen3.6 35B A3B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4/raw/e7759a52191808ad69737acd5d93f6f78c1af4f5/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Range-read safetensors headers found three files with tensor data totaling 25.025991648 GB: model.safetensors 22.443567616 GB, model_mtp.safetensors 1.689281536 GB, and model_visual.safetensors 0.893142496 GB. Stored tensor bytes split into BF16 6.682218208 GB, F8_E4M3 2.0381696 GB, F32 0.00024704 GB, and U8 16.3053568 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.426448896 GB. Auxiliary resident tensors, defined as model_visual plus model_mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Main routed expert tensors sum to 18.11963904 GB, or 0.07077984 GB per expert index. Fixed ordinary text traffic sums to 3.306809856 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from RedHatAI model card/API metadata, served config, base config comparison, safetensors index and direct shard header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately does not inherit NVIDIA's FP8-KV serving assumption because the RedHatAI config has kv_cache_scheme null and the model card does not request FP8 KV cache." }, { "id": "rippertnt--hyperclovax-seed-text-instruct-1-5b-q4-k-m-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF", "title": "HyperCLOVAX SEED Text Instruct 1.5B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the Q4_K_M GGUF artifact of HyperCLOVAX SEED Text Instruct 1.5B.", "model_family": "hyperclovax-seed-llama-dense", "base_model_proof": { "base_model": "naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B", "relation": "quantized", "source": "Hugging Face model card metadata, base-model API metadata, quantized repo config, and GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B. The base raw config is gated, so the audited architecture uses the quantized repo's own pinned config plus the selected GGUF header. The base-model API metadata still verifies the BF16 base parameter count and Llama architecture family." }, "architecture": { "canonical_architecture_id": "hyperclovax-seed-llama-1-5b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.81203968, "swept_params_b": 1.585547264, "auxiliary_resident_params_b": 0.226492416, "resident_weight_gb": 1.133974368, "swept_weight_gb": 1.002512384, "auxiliary_resident_weight_gb": 0.131461984, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for hyperclovax-seed-text-instruct-1.5b-q4_k_m.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q4_K_M linked file is 1.133974368 GB. Header tensor spans total 1.129914368 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.00406 GB. The main GGUF contains output.weight, token_embd.weight, blk.* tensors, and output_norm.weight. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 24, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The quantized repo config and GGUF header record a Llama-style 24-layer decoder with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected Q4_K_M GGUF artifact. The model card server example uses -c 2048, but the config and GGUF header both record a 131072-token maximum context." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6257999648219624, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and write traffic are outside Bounds Engine v1.", "notes": "The selected artifact uses a mixed Q4_K_M layout: tensor spans split into 0.77856768 GB Q4_K, 0.35094528 GB Q6_K, and 0.000401408 GB F32. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "HyperCLOVAX SEED GGUF HF API metadata", "url": "https://huggingface.co/api/models/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 3d2edd543d75372922aabd4ae538e6d181634b5d records a public non-gated GGUF repo with base_model naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B, region:us, 913990 downloads, GGUF architecture llama, 131072 context length, gguf.total 1812039680, and gguf.totalFileSize 1133974368." }, { "label": "HyperCLOVAX SEED GGUF model card", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card says this repo was converted to GGUF from naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B using llama.cpp via GGUF-my-repo, and shows the selected hyperclovax-seed-text-instruct-1.5b-q4_k_m.gguf artifact for llama-cli and llama-server use." }, { "label": "HyperCLOVAX SEED GGUF config", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF/raw/3d2edd543d75372922aabd4ae538e6d181634b5d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "tie_word_embeddings" ], "notes": "The quantized repo config records LlamaForCausalLM, model_type llama, BF16 source dtype, hidden size 2048, 24 layers, 16 attention heads, 8 KV heads, 128 head dimension, 7168 intermediate size, 110592 vocab size, tie_word_embeddings true, and 131072 max position embeddings." }, { "label": "HyperCLOVAX SEED base-model API metadata", "url": "https://huggingface.co/api/models/naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "license" ], "notes": "The base API metadata at commit 0728a47d632019a8da5f53b663db1c175dc04115 records a gated-auto Transformers Llama text-generation model with BF16 safetensors total 1585547264 parameters. The raw base config was gated during audit, so it is not used as the direct architecture source." }, { "label": "HyperCLOVAX SEED GGUF linked-object HEAD check", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF/tree/3d2edd543d75372922aabd4ae538e6d181634b5d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "The selected GGUF artifact resolves to x-repo-commit 3d2edd543d75372922aabd4ae538e6d181634b5d, x-linked-size 1133974368, x-xet-hash ba17dde49005bd8d17ff7fbe360bb1d973bc6b0ca4559fbfa91f55fa964be7a9, and final content-length 1133974368." }, { "label": "HyperCLOVAX SEED Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF/resolve/3d2edd543d75372922aabd4ae538e6d181634b5d/hyperclovax-seed-text-instruct-1.5b-q4_k_m.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 34 metadata entries and 219 tensors. The linked file is 1.133974368 GB. Tensor spans sum to 1.129914368 GB: output.weight 0.18579456 GB, output_norm.weight 0.000008192 GB, token_embd.weight 0.127401984 GB, and blk.* tensors 0.816709632 GB. Metadata/tokenizer/header/file overhead accounts for 0.00406 GB. Stored tensor bytes split into Q4_K 0.77856768 GB, Q6_K 0.35094528 GB, and F32 0.000401408 GB. The header records llama.block_count 24, context_length 131072, embedding_length 2048, feed_forward_length 7168, attention.head_count 16, attention.head_count_kv 8, and key/value head length 128." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from live HF API metadata, model card, quantized repo config, base-model API metadata, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the selected HyperCLOVAX SEED Text Instruct 1.5B Q4_K_M GGUF artifact. Do not infer the base repo raw config directly; access was gated during audit, so the architecture evidence is the quantized repo config plus GGUF header metadata." }, { "id": "sakamakismile--qwen3-6-27b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "sakamakismile/Qwen3.6-27B-NVFP4", "title": "SakamakiSmile Qwen3.6 27B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the SakamakiSmile compressed-tensors NVFP4 package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card, quantization recipe, and served config comparison", "config_compatible": true, "notes": "The live API does not expose base_model tags, but the model card identifies Qwen/Qwen3.6-27B as the original model and describes this repo as its NVFP4 quantized compressed-tensors release. Manual comparison found matching audited text and vision geometry fields against the pinned Qwen/Qwen3.6-27B base config: Qwen3_5ForConditionalGeneration, language_model_only false, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, resident vision tower, and 262144 max position embeddings. The target adds compressed-tensors NVFP4 quantization metadata while preserving the base architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.35672856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.732128496, "resident_weight_gb": 19.709536608, "swept_weight_gb": 16.245279616, "auxiliary_resident_weight_gb": 3.464256992, "resident_parameter_scope": "base logical Qwen3.6 parameters excluding absent MTP tensors, with exact stored compressed-tensors NVFP4/BF16/F8/F32 safetensors bytes", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from the single safetensors file because the package mixes packed U8 NVFP4 weights, F8_E4M3 scales, F32 global scales, and unquantized BF16 tensors. The model card says this older compressed-tensors release preserves the vision tower and has no MTP head; direct header grouping found no top-level MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers. The model card's production vLLM launch uses --kv-cache-dtype fp8, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The target preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and NVFP4 dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes a resident vision tower. This profile models ordinary text decode after any multimodal prefill, with no speculative MTP draft path because the package does not contain top-level MTP tensors." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8-attention-bf16-conv-f32-deltanet-state", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-deltanet-state", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-compressed-tensors-nvfp4-qwen3.6-multimodal-fp8-kv-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored NVFP4 packed weights, F8 scales, F32 global scales, unquantized BF16 tensors, FP8 attention KV, and fixed DeltaNet state. NVFP4 dequantization, activation traffic, compute overhead, state writes, and image/video prefill behavior are outside this memory-side bound.", "notes": "The config records compressed-tensors nvfp4-pack-quantized with 4-bit float weights and local dynamic 4-bit input activations, group_size 16, F8_E4M3 scale dtype, and kv_cache_scheme null. The model card recommends a 256K production vLLM launch with --kv-cache-dtype fp8; this profile follows that production launch for attention KV." }, "evidence": [ { "label": "SakamakiSmile Qwen3.6 27B NVFP4 API metadata and model card", "url": "https://huggingface.co/api/models/sakamakismile/Qwen3.6-27B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c, the API records a public non-gated safetensors repo with qwen3_5, 8-bit, compressed-tensors, and region:us tags. Current downloads are 169313. The API safetensors block reports F32 992, BF16 3006172400, F8_E4M3 1521909760, U8 12175278080, and total 16703361232. The pinned card identifies this as an Apache-2.0 NVFP4 quantized version of Qwen/Qwen3.6-27B, says the vision tower is preserved in BF16, says this older compressed-tensors release has no MTP head, and recommends production vLLM serving with 262144 max model length, max-num-seqs 2, and FP8 KV cache." }, { "label": "SakamakiSmile Qwen3.6 27B NVFP4 config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/raw/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The pinned config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors NVFP4 quantization, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config with depth 27 and hidden size 1152, and quantization ignores for model.visual*, lm_head, mlp.gate, and mlp.shared_expert_gate." }, { "label": "SakamakiSmile Qwen3.6 27B NVFP4 recipe", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/raw/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/recipe.yaml", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope", "auxiliary_resident_scope" ], "notes": "The recipe applies NVFP4 to Linear targets and ignores lm_head, all visual tensors, mlp.gate, and mlp.shared_expert_gate. This matches the card statement that the vision tower is preserved in BF16 and supports keeping visual tensors resident-only for ordinary text decode." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences across the audited top-level and text geometry fields between the target config and the pinned base config. The target vision_config adds an explicit bfloat16 dtype field; otherwise the checked vision geometry fields also match the base config. Layer type arrays are identical." }, { "label": "SakamakiSmile Qwen3.6 27B NVFP4 safetensors header", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/resolve/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The single safetensors header was range-read at repo SHA c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c. Stored tensor bytes total 19.709536608 GB: U8 12.175278080 GB, BF16 6.012344800 GB, F8_E4M3 1.521909760 GB, and F32 0.000003968 GB. The ordinary text swept subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 16.245279616 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual tensors, totals 3.464256992 GB. The header contains no top-level MTP tensors. lm_head.weight and model.language_model.embed_tokens.weight are separate BF16 tensors of shape 248320 x 5120 and 2.542796800 GB each." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, the model card, pinned compressed-tensors NVFP4 config and recipe, current base config comparison, direct range-read safetensors header, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted unquantized BF16 tensors, F8 scales, F32 scales, vision, embedding, and output-head storage. It is for ordinary text decode bounds after any multimodal prefill." }, { "id": "sakamakismile--qwen3-6-27b-text-nvfp4-mtp", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP", "title": "SakamakiSmile Qwen3.6 27B Text NVFP4 MTP", "summary": "Audited memory-side text-decode bounds profile for the text-only ModelOpt NVFP4 Qwen3.6 27B package with restored BF16 MTP tensors.", "model_family": "qwen3.6-dense-text-only", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata and served config comparison", "config_compatible": true, "notes": "The model card records Qwen/Qwen3.6-27B as the quantized base. Manual comparison found no differences across the audited text geometry fields between the target text_config and the pinned base text_config: 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, and 262144 native context. The target sets language_model_only true, removes the base visual tower, restores top-level BF16 MTP tensors, and adds ModelOpt NVFP4 quantization metadata." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b-text-only", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.320697856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.696097792, "resident_weight_gb": 19.6374752, "swept_weight_gb": 16.245279616, "auxiliary_resident_weight_gb": 3.392195584, "resident_parameter_scope": "base logical Qwen3.6 text-only parameters with exact stored ModelOpt NVFP4/BF16/F8/F32 safetensors bytes", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.language_model.embed_tokens.weight and top-level mtp tensors are resident for the package but not swept for ordinary non-speculative text decode; the base visual tower is physically absent from this text-only package", "notes": "Logical parameter counts follow the base Qwen3.6 27B text architecture with the visual tower removed. Bounds use exact stored bytes from the single safetensors header because the package mixes packed U8 NVFP4 weights, F8_E4M3 scales, F32 global scales, and unquantized BF16 tensors. The swept subset includes model.language_model tensors except input embeddings, plus lm_head.weight. The config records tie_word_embeddings false, so lm_head.weight is a separate output projection and remains swept for ordinary text decode. Top-level MTP tensors are resident-only for the ordinary non-speculative Bounds Engine v1 path." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers. The model card's recommended 256K production launch uses --kv-cache-dtype fp8, so attention K and V are charged as FP8 scalars." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The target preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, speculative MTP draft/verify traffic, and NVFP4 dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3_5ForConditionalGeneration is configured as language_model_only true for this package. This profile models ordinary text decode with speculative MTP disabled; the restored MTP tensors are resident but not swept on each ordinary generated token." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8-attention-bf16-conv-f32-deltanet-state", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8-attention-bf16-conv-f32-deltanet-state", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-qwen3.6-text-mtp-fp8-kv-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4 packed weights, F8 scales, F32 global scales, unquantized BF16 tensors, FP8 attention KV, and fixed DeltaNet state. NVFP4 dequantization, activation traffic, compute overhead, state writes, and speculative MTP draft/verify behavior are outside this memory-side bound.", "notes": "The config records ModelOpt NVFP4 with group size 16 and excludes lm_head, linear-attention conv1d weights, model.visual*, and top-level MTP tensors from quantization. hf_quant_config records kv_cache_quant_algo null, but the model card's recommended 256K production vLLM launch uses --kv-cache-dtype fp8 for concurrency headroom; this profile follows that production launch." }, "evidence": [ { "label": "SakamakiSmile Qwen3.6 27B Text NVFP4 MTP model card and API metadata", "url": "https://huggingface.co/api/models/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b" ], "notes": "At repo SHA 6f194695406a3bc88a00573187d5b2eecf984a99, the API records a public Apache-2.0 transformers text-generation repo with qwen3_5, qwen3.6, nvfp4, quantized, modelopt, MTP, speculative-decoding, blackwell, text-only, endpoints_compatible, 8-bit, region:us, and base_model:Qwen/Qwen3.6-27B metadata. Current downloads are 338680. The API safetensors block reports BF16 2970141696, F8_E4M3 1521909760, U8 12175278080, and total 16667329536. The card identifies this as a ModelOpt NVFP4 text-only sibling of Qwen/Qwen3.6-27B with BF16 MTP restored, says the visual tower was physically deleted, and recommends 256K vLLM serving with FP8 KV cache and qwen3_5_mtp speculative config." }, { "label": "SakamakiSmile Qwen3.6 27B Text NVFP4 MTP config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/raw/6f194695406a3bc88a00573187d5b2eecf984a99/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only true, no vision_config, tie_word_embeddings false, ModelOpt NVFP4 quantization, BF16 text config, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, MTP with one hidden layer, and quantization ignores for lm_head, linear-attention conv1d weights, model.visual*, and all top-level MTP modules." }, { "label": "SakamakiSmile Qwen3.6 27B Text NVFP4 MTP ModelOpt quantization config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/raw/6f194695406a3bc88a00573187d5b2eecf984a99/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "quantized_module_scope" ], "notes": "hf_quant_config.json records ModelOpt 0.43.0, quant_algo NVFP4, group_size 16, kv_cache_quant_algo null, and exclude_modules matching the served config's lm_head, linear-attention conv1d, and model.visual* exclusions." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences across 18 audited text geometry fields between the target text_config and the current base text_config: model type, hidden/intermediate sizes, layer count, attention head geometry, full_attention_interval, layer_types, max positions, tied embeddings, vocabulary, linear-attention key/value head counts and dimensions, convolution kernel, and mamba_ssm_dtype. The target differs intentionally by setting language_model_only true, removing vision_config, and adding ModelOpt NVFP4 quantization metadata." }, { "label": "SakamakiSmile Qwen3.6 27B Text NVFP4 MTP safetensors header", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/resolve/6f194695406a3bc88a00573187d5b2eecf984a99/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The single safetensors header was range-read at repo SHA 6f194695406a3bc88a00573187d5b2eecf984a99. Stored tensor bytes total 19.637475200 GB: U8 12.175278080 GB, BF16 5.940283392 GB, F8_E4M3 1.521909760 GB, and F32 0.000003968 GB. The ordinary text swept subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 16.245279616 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus top-level mtp tensors, totals 3.392195584 GB. lm_head.weight and model.language_model.embed_tokens.weight are separate BF16 tensors of shape 248320 x 5120 and 2.542796800 GB each. Top-level MTP tensors total 0.849398784 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned ModelOpt NVFP4 config and quantization config, current base text-config comparison, direct range-read safetensors header, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary non-speculative text decode bounds. It deliberately separates resident input-embedding/MTP weights from per-token swept language/logit weights and includes FP8 full-attention KV plus a fixed-state charge for DeltaNet linear-attention layers." }, { "id": "sgl-project--deepseek-v4-flash-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "sgl-project/DeepSeek-V4-Flash-FP8", "title": "SGL Project DeepSeek V4 Flash FP8", "summary": "Audited memory-side text-decode bounds profile for the SGL FP8 repackaging of DeepSeek V4 Flash.", "model_family": "deepseek-v4-flash-moe", "base_model_proof": { "base_model": "deepseek-ai/DeepSeek-V4-Flash", "relation": "quantized", "source": "SGL model card, served config comparison, official DeepSeek V4 Flash profile, and direct safetensors header grouping", "config_compatible": true, "notes": "The SGL model card states this is an FP8 repackaging of deepseek-ai/DeepSeek-V4-Flash with unchanged architecture, tokenizer, chat template, and reference encoding. The served config preserves the DeepSeek V4 Flash geometry and CSA/HCA compression fields while changing the packaging to all-FP8 e4m3 weights with dynamic activations." }, "architecture": { "canonical_architecture_id": "deepseek-v4-flash", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 294.03138004, "main_resident_weight_gb": 286.323482204, "auxiliary_resident_weight_gb": 7.707897836, "fixed_weight_gb": 9.23045846, "routed_expert_weight_gb": 1.082394624, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32_i64", "traffic_scope": "ordinary decode through layers 0-42 plus norm.weight, head.weight, and top-level hc_head tensors, excluding resident-only embed.weight and mtp.0 tensors", "auxiliary_scope": "embed.weight and mtp.0 tensors are resident for the checkpoint but not swept for each ordinary decode token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "The SGL safetensors index metadata.total_size matches the official mixed FP4/FP8 package and is not authoritative for this all-FP8 repack. Direct range-read shard headers and linked-object HEAD checks show the SGL package is about 294 GB. Routed expert tensors are byte-uniform across 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "sliding_window", "layers": 43, "kv_heads": 1, "head_dim": 512, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "All main layers keep a 128-token BF16 latent KV window in the official inference code." }, { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.00688, "notes": "Allocation coefficient for 21 ratio-4 main compressed BF16 caches, 20 ratio-128 main compressed BF16 caches, and 21 ratio-4 indexer BF16 caches." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.001504, "notes": "Read coefficient for full ratio-4 indexer scoring cache plus ratio-128 compressed cache at the default read context; the capped ratio-4 selected main-KV read is represented as a fixed read term." }, "notes": "DeepSeek V4 Flash uses Compressed Sparse Attention and Heavily Compressed Attention. Bounds Engine v1 linearizes the compressed-cache pieces from the official inference code." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.01220608, "read_gb_per_output_token": 0.011010048, "state_formula": "Compressor kv_state/score_state fixed buffers across 21 ratio-4 layers, 20 ratio-128 layers, and 21 ratio-4 indexer layers; fixed read term charges 512 selected compressed main-KV slots for the 21 ratio-4 indexer layers", "notes": "The allocation is true fixed compressor state. The read term is a default-workload cap approximation for the sparse top-k selected main-KV read because Bounds Engine v1 does not yet model min(index_topk, context/compression) directly." } ], "notes": "At the default 100k allocated context and 32k read context, this profile charges 0.705842176 GB allocation and 0.064774144 GB read traffic per output token for the BF16 window/compressed/indexer cache path." }, "notes": "The official inference forward path constructs MTP modules but does not call them for ordinary text decode; MTP tensors remain resident-only in this profile." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1, "kv_store_format": "bf16_window_plus_compressed_bf16_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "sparse_bf16_window_compressed_kv_plus_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "sglang-deepseek-v4-fp8-csa-hca-memory-bound", "dequantization_notes": "The memory-side bound charges exact stored FP8/BF16/F32/I64 safetensors bytes. FP8 dequantization, sparse-attention kernel efficiency, state writes, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records FP8 e4m3 dynamic quantization with 128x128 weight blocks and BF16 model dtype. KV/state traffic uses the already audited DeepSeek V4 Flash CSA/HCA BF16 cache path because the SGL card states architecture and reference encoding are unchanged from the base repo." }, "evidence": [ { "label": "SGL DeepSeek V4 Flash FP8 model card and API metadata", "url": "https://huggingface.co/api/models/sgl-project/DeepSeek-V4-Flash-FP8", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "weight_format", "total_params_b" ], "notes": "At commit ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17, the API reports a public non-gated MIT repo with safetensors, deepseek_v4, fp8, quantized, base_model deepseek-ai/DeepSeek-V4-Flash, region:us, 369947 downloads, and safetensors parameters BF16 2891654272, I64 2327040, F32 53747026, F8_E4M3 288014467072, total 290962195410." }, { "label": "SGL DeepSeek V4 Flash FP8 README", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8/raw/ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving" ], "notes": "The README states this is an FP8 repackaging of deepseek-ai/DeepSeek-V4-Flash and that model architecture, tokenizer, chat template, and reference encoding are unchanged from the base repo." }, { "label": "SGL DeepSeek V4 Flash FP8 config", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8/raw/ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17/config.json", "source_type": "config", "supports": [ "model_family", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "compressed_attention", "serving" ], "notes": "The config records DeepseekV4ForCausalLM, 43 hidden layers, one MTP layer, 256 routed experts, 6 experts per token, 1 shared expert, hidden size 4096, 64 attention heads, 1 KV head, head_dim 512, 1048576 max position embeddings, sliding_window 128, index_head_dim 128, index_topk 512, FP8 e4m3 dynamic quantization, 128x128 weight blocks, and compress_ratios with two uncompressed layers, 21 ratio-4 layers, 20 ratio-128 layers, and an uncalled MTP ratio entry." }, { "label": "SGL DeepSeek V4 Flash FP8 safetensors index and shard headers", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8/raw/ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format" ], "notes": "Range-read safetensors headers across all 46 shards found 69143 tensors totaling 294.031380040 GB: BF16 5.783308544 GB, I64 0.018616320 GB, F32 0.214988104 GB, and F8_E4M3 288.014467072 GB. Linked-object HEAD checks total 294.038841472 GB, leaving 0.007461432 GB of safetensors header/container overhead outside tensor payloads. Ordinary swept tensors under layers plus norm, head, and top-level hc_head tensors sum to 286.323482204 GB. Resident-only embed.weight plus mtp.0 tensors sum to 7.707897836 GB. Routed expert tensors sum to 277.093023744 GB, exactly 1.082394624 GB per expert index. Fixed ordinary-decode traffic including shared experts sums to 9.230458460 GB." }, { "label": "DeepSeek V4 Flash official profile and inference review", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/inference/model.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "mtp_resident_only" ], "notes": "The existing audited official DeepSeek V4 Flash profile reviewed the official inference code and found ordinary generation loops over the 43 main layers, applies norm and head, does not call self.mtp, keeps a 128-token BF16 window, allocates compressed BF16 cache slots according to compress_ratio, and uses ratio-4 indexer caches with index_topk 512. The SGL README states this architecture and reference encoding are unchanged." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, SGL README, served config, official DeepSeek V4 Flash profile/code review, safetensors index, linked-object HEAD checks, and direct range-read safetensors header byte grouping." }, "notes": "This all-FP8 SGL repack is expected to be resident_not_fit on 128GB local systems. It is substantially larger in resident bytes than the official mixed FP4/FP8 checkpoint and the DS4/GGUF local-oriented artifact." }, { "id": "speakleash--bielik-11b-v3-0-instruct-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "speakleash/Bielik-11B-v3.0-Instruct-awq", "title": "Bielik 11B v3.0 Instruct AWQ", "summary": "Audited memory-side text-decode bounds profile for the compressed-tensors AWQ package of Bielik 11B v3.0 Instruct.", "model_family": "bielik-llama-dense", "architecture": { "canonical_architecture_id": "bielik-11b-v3", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.168796672, "swept_params_b": 11.037200384, "auxiliary_resident_params_b": 0.131596288, "resident_weight_gb": 6.192805344, "swept_weight_gb": 5.929612768, "auxiliary_resident_weight_gb": 0.263192576, "resident_parameter_scope": "base BF16 API logical parameters with direct compressed-tensors safetensors stored-byte totals", "swept_parameter_scope": "model.layers plus model.norm plus separate lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens.weight is resident for the package but not swept as a full matrix for ordinary text decode", "notes": "Bounds use exact stored bytes from safetensors headers because the compressed-tensors package mixes packed I32 weight_packed tensors, BF16 scale/zero-point/ignored-module tensors, and tiny I64 shape tensors. The base repo API records 11.168796672B BF16 parameters; the base raw config is gated in this audit environment, so this profile uses the served quantized config directly. The config records tie_word_embeddings false, and the headers contain a separate BF16 lm_head.weight." }, "kv_adapter": { "kind": "full_context", "layers": 50, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 50 Llama decoder layers, 32 attention heads, 8 KV heads, 128 head dimension, no sliding-window setting, and 32768 max positions, so this profile charges full-context BF16 K and V streams for cached text decode." }, "notes": "Bielik 11B v3.0 Instruct AWQ uses LlamaForCausalLM. This profile models ordinary cached text decode only; prefill, tokenizer overhead, and quantization quality are outside Bounds Engine v1." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.5544738189679169, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "compressed-tensors-awq-int4-llama-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors bytes: packed I32 weights, BF16 scales, BF16 zero points, BF16 ignored modules, and tiny I64 shape tensors from safetensors headers. Dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors pack-quantized 4-bit integer weights with group_size 128, asymmetric quantization, BF16 model dtype, and kv_cache_scheme null. The model card metadata includes a stale gguf tag, but the served artifacts audited here are safetensors." }, "evidence": [ { "label": "Bielik 11B v3.0 Instruct AWQ model card and API metadata", "url": "https://huggingface.co/api/models/speakleash/Bielik-11B-v3.0-Instruct-awq", "source_type": "model_card", "supports": [ "repo", "base_model", "license", "pipeline", "revision", "downloads", "serving", "storage_accounting" ], "notes": "At repo SHA 0de6dd10ddcbd570b3844f1ae0e98f8e4800b548, the API records a public Apache-2.0 text-generation repo derived from speakleash/Bielik-11B-v3.0-Instruct, with transformers, safetensors, llama, finetuned, multilingual, text-generation-inference, compressed-tensors, and region:us tags. Current downloads are 628428. The API safetensors block reports BF16: 348803072, I32: 10990387200, I64: 700. The model card says the repo contains AWQ files for Bielik-11B-v3.0-Instruct and warns that quantized models can reduce response quality." }, { "label": "Bielik 11B v3.0 Instruct AWQ config", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct-awq/raw/0de6dd10ddcbd570b3844f1ae0e98f8e4800b548/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, BF16 dtype, hidden size 4096, intermediate size 14336, 50 hidden layers, 32 attention heads, 8 KV heads, 128 head dimension, 32768 max position embeddings, rope_theta 1000000, vocab size 32128, tie_word_embeddings false, and compressed-tensors AWQ with group size 128, 4-bit integer weights, asymmetric quantization, minmax observer, lm_head ignored, and kv_cache_scheme null." }, { "label": "Bielik 11B v3.0 Instruct base API metadata", "url": "https://huggingface.co/api/models/speakleash/Bielik-11B-v3.0-Instruct", "source_type": "derived_calculation", "supports": [ "resident_params_b", "base_model" ], "notes": "The base repo API records commit 735bfee1125fe8b497ac2769de94822a11f77167, Apache-2.0 license metadata, gated:auto, and BF16 safetensors parameters 11168796672. Raw base config access returned 401 Unauthorized in this audit environment, so architecture geometry is taken from the served quantized config rather than an unaudited base comparison." }, { "label": "Bielik 11B v3.0 Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct-awq/resolve/0de6dd10ddcbd570b3844f1ae0e98f8e4800b548/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The index records total_size 6192805344 bytes across two shards. Direct range-read safetensors headers match the index exactly and found 1503 tensors totaling 6.192805344 GB: I32 5.495193600 GB, BF16 0.697606144 GB, and I64 0.000005600 GB. The swept ordinary text subset, defined as model.layers plus model.norm.weight plus lm_head.weight, totals 5.929612768 GB. The resident-only model.embed_tokens.weight tensor is 0.263192576 GB. Header buckets are MLP 4.576053600 GB, self-attention 1.089539200 GB, lm_head 0.263192576 GB, input embedding 0.263192576 GB, layer norms 0.000819200 GB, and final norm 0.000008192 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned served compressed-tensors config, base API metadata, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped row, which mislabeled the artifact as GGUF and undercounted AWQ side tensors plus the unquantized output head and input embeddings." }, { "id": "stelterlab--mistral-small-24b-instruct-2501-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stelterlab/Mistral-Small-24B-Instruct-2501-AWQ", "title": "Mistral Small 24B Instruct 2501 AWQ", "summary": "Audited memory-side text-decode bounds profile for the stelterlab AWQ Int4 package of Mistral Small 24B Instruct 2501.", "model_family": "mistral-small-dense-awq", "base_model_proof": { "base_model": "mistralai/Mistral-Small-24B-Instruct-2501", "relation": "quantized", "source": "Hugging Face API metadata, AWQ model card, served AWQ config, current base config comparison, and direct safetensors shard headers", "config_compatible": true, "notes": "The AWQ repo API and card identify mistralai/Mistral-Small-24B-Instruct-2501 as the quantized base model. Manual comparison against the current base config found matching checked text geometry and context fields; the AWQ artifact adds AWQ quantization metadata and changes torch_dtype to float16." }, "architecture": { "canonical_architecture_id": "mistral-small-24b-2501", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 23.5724032, "swept_params_b": 22.90131456, "auxiliary_resident_params_b": 0.67108864, "resident_weight_gb": 14.2342656, "swept_weight_gb": 12.89208832, "auxiliary_resident_weight_gb": 1.34217728, "resident_parameter_scope": "logical AWQ model parameters represented by the safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records sliding_window null, so this profile charges full-context K and V streams for all 40 language layers." }, "notes": "Dense MistralForCausalLM profile using the served AWQ config and direct safetensors header grouping." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6038529665061898, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Mistral Small 24B Instruct AWQ API metadata", "url": "https://huggingface.co/api/models/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit cbda099649a0188dd888d44f0e4964d8d982dc9a, the API reports a public non-gated vLLM text-generation repo with base_model mistralai/Mistral-Small-24B-Instruct-2501, Apache-2.0 license, 4-bit and awq tags, region:us, current downloads 235342, and safetensors logical parameters I32: 22229811200, F16: 1342592000, total: 23572403200." }, { "label": "Mistral Small 24B Instruct AWQ served config", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ/raw/cbda099649a0188dd888d44f0e4964d8d982dc9a/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records MistralForCausalLM with hidden_size 5120, intermediate_size 32768, 40 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, sliding_window null, tie_word_embeddings false, vocab_size 131072, torch_dtype float16, rope_theta 100000000, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Mistral Small 24B Instruct AWQ model card", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ/blob/cbda099649a0188dd888d44f0e4964d8d982dc9a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "context_notes" ], "notes": "The model card states the package is INT4 GEMM AWQ quantization done with AutoAWQ, identifies the original model as Mistral-Small-24B-Instruct-2501, and describes a 32k context window." }, { "label": "Mistral Small 24B Instruct base config", "url": "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501/raw/9527884be6e5616bdd54de542f9ae13384489724/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The current base config records the same checked architecture fields as the AWQ config after excluding quantization_config, _name_or_path, torch_dtype, transformers_version, and use_cache: MistralForCausalLM, hidden_size 5120, intermediate_size 32768, 40 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, sliding_window null, tie_word_embeddings false, vocab_size 131072, and rope_theta 100000000." }, { "label": "Mistral Small 24B Instruct AWQ safetensors index and shard headers", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ/raw/cbda099649a0188dd888d44f0e4964d8d982dc9a/model.safetensors.index.json", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The safetensors index metadata reports total_size 14234265600 bytes across three shards, matching direct range-read tensor spans. Headers contain 923 tensors totaling 14.234265600 GB: 11.201740800 GB I32 tensors and 3.032524800 GB F16 tensors. Stored suffix totals are qweight 11.114905600 GB, qzeros 0.086835200 GB, scales 0.347340800 GB, and F16 weight tensors 2.685184000 GB. model.embed_tokens.weight and lm_head.weight each have shape [131072, 5120] and contribute 1.342177280 GB. Linked shard sizes sum to 14.234370648 GB, or 105048 bytes of safetensors header overhead above tensor payload." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served AWQ config, model card, current base config comparison, params.json, safetensors index metadata, linked-object HEAD checks, and direct shard header byte grouping." }, "notes": "This profile supersedes the scraped flat 0.5 byte/parameter estimate by using exact stored AWQ tensor bytes and by excluding only the input embedding matrix from ordinary text-decode swept traffic." }, { "id": "stepfun-ai--step-3-5-flash", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stepfun-ai/Step-3.5-Flash", "title": "StepFun Step 3.5 Flash BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the BF16/F32 Step 3.5 Flash sparse MoE language model.", "model_family": "step3p5-moe", "architecture": { "canonical_architecture_id": "step-3-5-flash", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 398.768626944, "main_resident_weight_gb": 392.856368896, "auxiliary_resident_weight_gb": 5.912258048, "fixed_weight_gb": 12.349110016, "routed_expert_weight_gb": 1.32120576, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32_with_indexed_mtp", "traffic_scope": "ordinary text decode through model.layers.0-44, model.norm, and lm_head, with fused routed expert tensors divided into uniform expert-index groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight and indexed model.layers.45-47 MTP tensors are resident for the checkpoint but not swept as full matrices for each ordinary target-model decode token", "shared_expert_notes": "The config records share_expert_dim 1280 for MoE layers. Shared expert tensors are included in fixed_weight_gb because they are always read when a MoE layer runs.", "notes": "Layers 3-44 are MoE layers. The checkpoint stores routed experts as fused per-layer BF16 tensors named moe.down_proj, moe.gate_proj, and moe.up_proj with expert dimension 288, so routed_expert_weight_gb is the grouped routed tensor byte count divided by 288 expert indexes. The small F32 router_bias tensors are included in fixed ordinary decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "Among ordinary model.layers.0-44, layer_types mark layers 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, and 44 as full_attention. These layers use 64 query heads, 8 KV groups, and 128 head dimension." }, { "kind": "sliding_window", "layers": 33, "kv_heads": 8, "head_dim": 128, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 33 ordinary target-model layers are sliding_attention layers with config.sliding_window 512, 96 query heads, 8 KV groups, and 128 head dimension. The config has three additional layer_types entries for indexed MTP layers 45-47, which are outside this ordinary decode profile." } ], "notes": "The config does not record a quantized KV cache scheme for the BF16 checkpoint, so Bounds Engine v1 charges BF16 full/sliding K/V cache. Speculative MTP throughput is outside this ordinary target-model text-decode profile." }, "notes": "Step3p5ForCausalLM wraps a Step3p5Model plus lm_head. The remote forward path iterates self.layers[:config.num_hidden_layers], so indexed MTP tensors in model.layers.45-47 are resident package tensors but not ordinary target-model decode layers." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000001213335246, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-bf16-step3p5-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored BF16/F32 safetensors payload bytes and BF16 hybrid full/sliding KV cache traffic. Activation traffic, router compute, expert compute, expert-parallel communication, cache writes, and speculative MTP decode are outside this memory-side bound.", "notes": "The config records bfloat16 dtype and the API safetensors block records 199384289280 BF16 parameters plus 12096 F32 parameters. The model card shows BF16 vLLM/SGLang serving examples without an explicit quantized KV cache requirement." }, "evidence": [ { "label": "Step 3.5 Flash API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.5-Flash", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit ab446a3de5e171ea341227e24bb1f090e1b771f7, the live API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, custom_code, step3p5, eval-results, and region:us tags. Current downloads are 178273. The API safetensors block records BF16 199384289280, F32 12096, total 199384301376." }, { "label": "Step 3.5 Flash model card", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/raw/ab446a3de5e171ea341227e24bb1f090e1b771f7/README.md", "source_type": "model_card", "supports": [ "architecture", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "kv_adapter", "serving" ], "notes": "The model card describes Step 3.5 Flash as a sparse MoE model with a 196B-parameter package, about 11B activated parameters per token, 256K context, a 3:1 sliding-window/full-attention ratio, MTP-3, and BF16 vLLM/SGLang deployment examples." }, { "label": "Step 3.5 Flash config", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/raw/ab446a3de5e171ea341227e24bb1f090e1b771f7/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The config records Step3p5ForCausalLM, step3p5, BF16, 45 ordinary hidden layers, 262144 max context, 12 full_attention entries among ordinary layers, 33 sliding_attention ordinary layers, sliding_window 512, 64 full-attention query heads, 96 sliding-attention query heads, 8 KV groups, head_dim 128, 288 experts, top-k 8, MoE layers 3-44, share_expert_dim 1280, and 3 next-token prediction layers." }, { "label": "Step 3.5 Flash remote modeling code", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/raw/ab446a3de5e171ea341227e24bb1f090e1b771f7/modeling_step3p5.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "routed_experts", "shared_experts_per_token", "auxiliary_resident_scope" ], "notes": "Manual review found Step3p5MoEMLP uses fused MoELinear tensors with shape [num_experts, out_features, in_features], router top-k selection, and separate shared expert modules in MoE layers. Step3p5Attention selects sliding-window attention from config.layer_types and config.sliding_window. Step3p5Model.forward iterates self.layers[:config.num_hidden_layers], excluding indexed layers 45-47 from ordinary target-model decode." }, { "label": "Step 3.5 Flash safetensors index and shard headers", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/resolve/ab446a3de5e171ea341227e24bb1f090e1b771f7/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 398768626944 bytes across 44 referenced files. Direct range-read safetensors headers found 804 tensors with payload bytes 398.768626944 GB and header/container overhead 0.000093256 GB. Payload dtype split is BF16 398.768578560 GB and F32 0.000048384 GB. Resident-only input embedding and MTP tensors sum to 5.912258048 GB. Main ordinary text resident bytes sum to 392.856368896 GB. Fused routed expert tensors sum to 380.507258880 GB, or 1.321205760 GB per expert index across 288 experts. Fixed ordinary text traffic, including self-attention, dense layers, router gates/biases, shared experts, norms, and lm_head, sums to 12.349110016 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, served config, remote modeling code, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated dense estimate. It models the BF16/F32 Step 3.5 Flash checkpoint as a sparse MoE language model with exact fused expert byte groups and BF16 hybrid full/sliding KV cache." }, { "id": "stepfun-ai--step-3-7-flash-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stepfun-ai/Step-3.7-Flash-NVFP4", "title": "StepFun Step 3.7 Flash NVFP4", "summary": "Audited memory-side text-decode bounds profile for the NVFP4 Step 3.7 Flash multimodal MoE package.", "model_family": "step3p7-moe-vlm", "base_model_proof": { "base_model": "stepfun-ai/Step-3.7-Flash", "relation": "quantized", "source": "StepFun model card, served NVFP4 config, BF16 Step 3.7 Flash config comparison, ModelOpt quantization sidecar, and safetensors header review", "config_compatible": true, "notes": "The model card presents this repo as the NVFP4 variant of Step 3.7 Flash. Manual comparison against the BF16 Step 3.7 Flash config found matching checked text, MoE, attention, MTP, and vision geometry. The NVFP4 config adds ModelOpt quantization metadata, explicit top-level duplicate fields, and explicit defaults such as partial_rotary_factor; the executable text_config and vision_config geometry used by this profile is preserved." }, "architecture": { "canonical_architecture_id": "step-3-7-flash", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 129.241354856, "main_resident_weight_gb": 119.36706724, "auxiliary_resident_weight_gb": 9.874287616, "fixed_weight_gb": 12.349110376, "routed_expert_weight_gb": 0.371590128, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_modelopt_nvfp4_u8_f8_bf16_f32_with_indexed_mtp", "traffic_scope": "ordinary text decode through model.language_model and lm_head, with fused expert tensors divided into uniform expert-index groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.language_model.embed_tokens.weight, model.vision_model, model.vit_large_projector, and indexed model.layers.45-47 MTP tensors are resident for the package but not swept for each ordinary decode token", "shared_expert_notes": "The text config records one shared expert per MoE layer via share_expert_dim 1280. Shared expert tensors are unquantized BF16 and are included in fixed_weight_gb because they are always read when a MoE layer runs.", "notes": "Layers 3-44 are MoE layers. The checkpoint stores routed experts as fused per-layer tensors, not per-expert tensor names: moe.down_proj/gate_proj/up_proj U8 weights, F8_E4M3 weight scales, and small F32 input/weight scale tensors. The routed expert total is the sum of those fused tensors across 42 MoE layers divided by 288 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text_config layer_types mark layers 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, and 44 as full_attention. These layers use 64 query heads, 8 KV groups, and 128 head dimension." }, { "kind": "sliding_window", "layers": 33, "kv_heads": 8, "head_dim": 128, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 33 text layers are sliding_attention layers. The custom attention code enables sliding-window attention for these layers and uses text_config.sliding_window 512 with 96 query heads, 8 KV groups, and 128 head dimension." } ], "notes": "The ModelOpt sidecar records FP8 KV cache quantization and the model card vLLM/SGLang commands request FP8 KV cache for the NVFP4 model. Vision prefill and speculative MTP throughput are outside this ordinary text-decode profile." }, "notes": "Step3p7ForConditionalGeneration wraps a StepRobotics vision encoder, multimodal projector, Step3p5 MoE language model, and indexed MTP tensors. This profile models ordinary cached text decode after any multimodal prefill, with speculative MTP disabled for the throughput bound." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8_e4m3", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8_e4m3", "kv_read_bytes_per_scalar": 1, "runtime_format": "modelopt-nvfp4-fp8-kv-step3p7-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4 U8 payloads, F8_E4M3 scales, BF16/F32 excluded tensors, FP8 KV cache traffic, and the exact safetensors tensor spans. NVFP4 dequantization, activation traffic, expert-parallel communication, attention kernels, and cache writes are outside this memory-side bound.", "notes": "The ModelOpt sidecar records quant_algo NVFP4, kv_cache_quant_algo FP8, and group size 16. The config quantization_config records 4-bit float weights and activations, FP8 KV cache scheme, and ModelOpt quantization method." }, "evidence": [ { "label": "Step 3.7 Flash NVFP4 API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.7-Flash-NVFP4", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 4275532ffd9a9496ff36b7a2dc4a9db1048da438, the live API records a public non-gated Apache-2.0 image-text-to-text repo with transformers, custom_code, step3p7, vision-language, multimodal, moe, modelopt, 8-bit, and region:us tags. Current downloads are 241770. The API safetensors block records F32 84672, BF16 11111674624, F8_E4M3 11890851840, U8 95126814720, total 103810330432." }, { "label": "Step 3.7 Flash model card", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "architecture", "serving", "kv_adapter" ], "notes": "The model card describes Step 3.7 Flash as a 198B sparse MoE VLM with a 196B language backbone, 1.8B vision encoder, about 11B activated parameters per token, and 256k context. It documents NVFP4 vLLM and SGLang launch commands using ModelOpt quantization, expert parallelism, and FP8 KV cache." }, { "label": "Step 3.7 Flash NVFP4 config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The served config records Step3p7ForConditionalGeneration, a 45-layer Step3p5ForCausalLM text_config, 262144 max context, 12 full_attention layers, 33 sliding_attention layers, sliding_window 512, 64 full-attention query heads, 96 sliding-attention query heads, 8 KV groups, head_dim 128, 288 experts, top-k 8, MoE layers 3-44, share_expert_dim 1280, 3 next-token prediction layers, BF16 runtime dtype, and a 47-layer 1536-wide vision encoder." }, { "label": "Step 3.7 Flash NVFP4 ModelOpt quantization sidecar", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_adapter", "serving" ], "notes": "The sidecar records producer modelopt 0.45.0.dev37, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and exclusions for lm_head, early dense layers, self-attention, router gate, shared experts, vision model, projector, and layers 45-47 MTP tensors." }, { "label": "Step 3.7 Flash BF16 config comparison", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison against the BF16 Step 3.7 Flash config checked 46 top-level, text_config, and vision_config geometry fields. The checked executable geometry matches; observed differences are explicit top-level duplicate fields or explicit defaults in the NVFP4 config, plus the NVFP4 quantization_config." }, { "label": "Step 3.7 Flash remote modeling code", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/modeling_step3p7.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "routed_experts", "shared_experts_per_token", "auxiliary_resident_scope" ], "notes": "Manual review found Step3p7MoEMLP uses fused MoELinear tensors with shape [num_experts, out_features, in_features], router top-k selection, and a separate shared expert in MoE layers. Step3p7DecoderLayer instantiates MoE plus shared expert only for configured MoE layers and uses sliding-window attention when layer_types marks a layer as sliding_attention. The model class ignores model.layers.45-47 unexpected keys in the ordinary base class, matching the separate indexed MTP tensors." }, { "label": "Step 3.7 Flash NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/resolve/4275532ffd9a9496ff36b7a2dc4a9db1048da438/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 129241604232 bytes and total_parameters 103810330432 across 14 referenced files, including model-mtp-bf16.safetensors. Direct range-read safetensors headers found 1939 tensors with payload bytes 129.241354856 GB and header/container overhead 0.000249376 GB. Payload dtype split is U8 95.126814720 GB, BF16 22.223349248 GB, F8_E4M3 11.890851840 GB, and F32 0.000339048 GB. Resident-only input embedding, vision/projector, and MTP tensors sum to 9.874287616 GB. Main ordinary text resident bytes sum to 119.367067240 GB. Fused routed expert tensors sum to 107.017956864 GB, or 0.371590128 GB per expert index across 288 experts. Fixed ordinary text traffic, including self-attention, dense layers, router gates, shared experts, norms, and lm_head, sums to 12.349110376 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live Hugging Face API metadata, model card, served config, ModelOpt quantization sidecar, BF16 config comparison, remote modeling code, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile supersedes the generated dense estimate. It models the NVFP4 checkpoint as a sparse MoE VLM with exact fused expert byte groups and FP8 hybrid full/sliding KV cache." }, { "id": "stepfun-ai--step-3-7-flash", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stepfun-ai/Step-3.7-Flash", "title": "StepFun Step 3.7 Flash BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the BF16/F32 Step 3.7 Flash sparse MoE multimodal package.", "model_family": "step3p7-moe-vlm", "architecture": { "canonical_architecture_id": "step-3-7-flash", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 402.730656512, "main_resident_weight_gb": 392.856368896, "auxiliary_resident_weight_gb": 9.874287616, "fixed_weight_gb": 12.349110016, "routed_expert_weight_gb": 1.32120576, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32_with_vision_projector_and_indexed_mtp", "traffic_scope": "ordinary text decode through model.language_model layers 0-44 and lm_head, with fused routed expert tensors divided into uniform expert-index groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.language_model.embed_tokens.weight, model.vision_model, model.vit_large_projector, and indexed model.layers.45-47 MTP tensors are resident for the package but not swept as full matrices for each ordinary target-model decode token", "shared_expert_notes": "The text config records one shared expert per MoE layer via share_expert_dim 1280. Shared expert tensors are included in fixed_weight_gb because they are always read when a MoE layer runs.", "notes": "Layers 3-44 are MoE layers. The checkpoint stores routed experts as fused per-layer BF16 tensors named moe.down_proj, moe.gate_proj, and moe.up_proj with expert dimension 288, so routed_expert_weight_gb is the grouped routed tensor byte count divided by 288 expert indexes. Small F32 router_bias tensors are included in fixed ordinary decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The text_config layer_types mark layers 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, and 44 as full_attention. These layers use 64 query heads, 8 KV groups, and 128 head dimension." }, { "kind": "sliding_window", "layers": 33, "kv_heads": 8, "head_dim": 128, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The remaining 33 ordinary text layers are sliding_attention layers. The custom attention code enables sliding-window attention for these layers and uses text_config.sliding_window 512 with 96 query heads, 8 KV groups, and 128 head dimension." } ], "notes": "The BF16 checkpoint config and BF16 model-card launch examples do not record an explicit quantized KV cache requirement, so Bounds Engine v1 charges BF16 full/sliding K/V cache. Vision prefill and speculative MTP throughput are outside this ordinary text-decode profile." }, "notes": "Step3p7ForConditionalGeneration wraps a StepRobotics vision encoder, multimodal projector/downsamplers, Step3p5 MoE language model, and indexed MTP tensors. This profile models ordinary cached text decode after any multimodal prefill, with speculative MTP disabled for the throughput bound." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000120139856, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-bf16-step3p7-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges exact stored BF16/F32 safetensors payload bytes and BF16 hybrid full/sliding KV cache traffic. Activation traffic, router compute, expert compute, expert-parallel communication, cache writes, vision prefill, and speculative MTP decode are outside this memory-side bound.", "notes": "The config records bfloat16 text dtype and the API safetensors block records BF16 201365304064, F32 12096, total 201365316160. The model card shows BF16 vLLM/SGLang serving examples without an explicit quantized KV cache flag." }, "evidence": [ { "label": "Step 3.7 Flash API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.7-Flash", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit 5f6244077ac62e04eec3f320501ff8c2b293373a, the live API records a public non-gated Apache-2.0 image-text-to-text repo with transformers, safetensors, custom_code, step3p7, vision-language, multimodal, moe, eval-results, and region:us tags. Current downloads are 146322. The API safetensors block records BF16 201365304064, F32 12096, total 201365316160." }, { "label": "Step 3.7 Flash model card", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/README.md", "source_type": "model_card", "supports": [ "architecture", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The model card describes Step 3.7 Flash as a 198B sparse MoE VLM with a 196B language backbone, 1.8B vision encoder, about 11B activated parameters per token, and 256k context. It documents BF16 vLLM and SGLang launch commands and separate NVFP4 examples that require FP8 KV cache." }, { "label": "Step 3.7 Flash config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The served config records Step3p7ForConditionalGeneration, a 45-layer Step3p5ForCausalLM text_config, 262144 max context, 12 full_attention layers, 33 sliding_attention layers, sliding_window 512, 64 full-attention query heads, 96 sliding-attention query heads, 8 KV groups, head_dim 128, 288 experts, top-k 8, MoE layers 3-44, share_expert_dim 1280, 3 next-token prediction layers, BF16 text dtype, and a 47-layer 1536-wide vision encoder." }, { "label": "Step 3.7 Flash remote modeling code", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/modeling_step3p7.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "routed_experts", "shared_experts_per_token", "auxiliary_resident_scope" ], "notes": "Manual review found Step3p7MoEMLP uses fused MoELinear tensors with shape [num_experts, out_features, in_features], router top-k selection, and a separate shared expert in MoE layers. Step3p7DecoderLayer instantiates MoE plus shared expert only for configured MoE layers and uses sliding-window attention when layer_types marks a layer as sliding_attention. The model class ignores model.layers.45-47 unexpected keys in the ordinary base class, matching the separate indexed MTP tensors." }, { "label": "Step 3.7 Flash safetensors index and shard headers", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/resolve/5f6244077ac62e04eec3f320501ff8c2b293373a/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 402730656512 bytes across 26 referenced files. Direct range-read safetensors headers found 1471 tensors with payload bytes 402.730656512 GB and header/container overhead 0.000177120 GB. Payload dtype split is BF16 402.730608128 GB and F32 0.000048384 GB. Resident-only input embedding, vision/projector, and MTP tensors sum to 9.874287616 GB. Main ordinary text resident bytes sum to 392.856368896 GB. Fused routed expert tensors sum to 380.507258880 GB, or 1.321205760 GB per expert index across 288 experts. Fixed ordinary text traffic, including self-attention, dense layers, router gates/biases, shared experts, norms, and lm_head, sums to 12.349110016 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, model card, served config, remote modeling code, safetensors index, linked shard HEAD checks, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated dense estimate. It models the BF16/F32 Step 3.7 Flash checkpoint as a sparse MoE VLM with exact fused expert byte groups and BF16 hybrid full/sliding KV cache." }, { "id": "stepfun-ai--step3-vl-10b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stepfun-ai/Step3-VL-10B", "title": "StepFun Step3-VL 10B BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Step3-VL 10B multimodal package.", "model_family": "step3-vl-qwen3-dense", "base_model_proof": { "base_model": "stepfun-ai/Step3-VL-10B-Base", "relation": "finetune", "source": "Hugging Face model metadata and direct base config comparison", "config_compatible": true, "notes": "The repo metadata records stepfun-ai/Step3-VL-10B-Base as the finetune base model. Manual comparison found matching checked architecture fields between this chat config and the base config except max_position_embeddings: the chat repo records 65536 while the base repo records 40960. This profile uses the served chat repo config as authoritative for max context." }, "architecture": { "canonical_architecture_id": "step3-vl-10b", "max_context_tokens": 65536, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.171750144, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 2.60334464, "resident_weight_gb": 20.343500288, "swept_weight_gb": 15.136811008, "auxiliary_resident_weight_gb": 5.20668928, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "model.layers, model.norm, and lm_head safetensors headers", "auxiliary_scope": "model.embed_tokens, vision_model, and vit_large_projector tensors are resident for the package but not swept as full matrices for ordinary cached text decode", "notes": "Bounds use exact BF16 stored bytes from safetensors headers. The custom checkpoint remaps the language decoder to model.layers/model.norm/model.embed_tokens and stores lm_head separately. Ordinary text decode sweeps transformer layers, final norm, and lm_head; input embedding, vision encoder, and visual projector are resident-only for decode after multimodal prefill." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served text_config is Qwen3ForCausalLM with sliding_window null and use_sliding_window false, so the v1 profile charges full-context BF16 K and V streams for all 36 decoder layers." }, "notes": "StepVLForConditionalGeneration wraps a StepRoboticsVisionEncoder plus a Qwen3Model language decoder. This profile models ordinary cached text decode after image features have been embedded; vision encoder prefill throughput is outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "step3-vl-qwen3-bf16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Vision encoder compute, image prefill, projector compute, attention kernels, activation traffic, and cache writes are outside this memory-side bound.", "notes": "The model card recommends BF16 inference and documents vLLM and SGLang serving paths. KV cache is charged as BF16 because the served config does not declare a KV-cache quantization scheme." }, "evidence": [ { "label": "Step3-VL 10B model card and API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step3-VL-10B", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 5026053b0c2f5dfaa08fc2d149384162c3c8bca1, the API records a public Apache-2.0 image-text-to-text repo derived from stepfun-ai/Step3-VL-10B-Base, with custom_code, step_robotics, safetensors, arxiv:2601.09668, and region:us tags. Current downloads are 351926. The API safetensors block reports BF16 10171750144 parameters. The card describes a 10B multimodal model with a 1.8B perception encoder, Qwen3-8B decoder, 64K max sequence during PaCoRe training, BF16 inference, and vLLM/SGLang serving commands." }, { "label": "Step3-VL 10B config", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/raw/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "auxiliary_resident_scope" ], "notes": "The served config records StepVLForConditionalGeneration, step_robotics, image token id 151679, vision encoder width 1536 with 47 layers and 16 heads, and a Qwen3ForCausalLM text_config with BF16 dtype, 36 layers, hidden size 4096, intermediate size 12288, 32 attention heads, 8 KV heads, head_dim 128, tie_word_embeddings false, max_position_embeddings 65536, sliding_window null, use_sliding_window false, and vocab_size 151936." }, { "label": "Step3-VL 10B base config comparison", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B-Base/raw/b84b55f58f5b65d108ef5993f4b71e11d421159c/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison of the pinned chat config and current base config found matching checked top-level, vision, and text architecture fields. The only checked difference is max_position_embeddings: 65536 in the chat config versus 40960 in the base config." }, { "label": "Step3-VL 10B remote modeling code", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/raw/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/modeling_step_vl.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "auxiliary_resident_scope" ], "notes": "Manual review found StepRoboticsModel constructs StepRoboticsVisionEncoder, Qwen3Model(config.text_config), and vit_large_projector, then passes attention_mask, past_key_values, use_cache, and cache_position through to the Qwen3 language model. The checkpoint conversion mapping stores language decoder tensors under model.layers/model.norm/model.embed_tokens and top-level lm_head while vision tensors remain under vision_model and the projector under vit_large_projector." }, { "label": "Step3-VL 10B safetensors index and shard headers", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/resolve/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout", "dtype_split" ], "notes": "The index maps 1066 tensors across five shards. Direct range-read safetensors headers found BF16 tensor payloads totaling 20.343500288 GB. Ordinary text swept tensors are model.layers 13.892143104 GB, model.norm 0.000008192 GB, and lm_head.weight 1.244659712 GB, totaling 15.136811008 GB. Resident-only tensors are model.embed_tokens.weight 1.244659712 GB, vision_model 3.911697920 GB, and vit_large_projector.weight 0.050331648 GB, totaling 5.206689280 GB. model.embed_tokens.weight and lm_head.weight are separate BF16 tensors with shape [151936, 4096]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, the model card, pinned config, base config comparison, remote modeling code, safetensors index, and direct range-read safetensors shard headers." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the whole BF16 resident package as swept language traffic and did not separate vision/projector/input-embedding resident-only bytes." }, { "id": "stepfun-ai--step3", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "stepfun-ai/step3", "title": "StepFun Step3 BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 Step3 sparse MoE multimodal checkpoint.", "model_family": "step3-vl-moe", "architecture": { "canonical_architecture_id": "step3", "max_context_tokens": 65536, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 641.939825664, "main_resident_weight_gb": 630.74238464, "auxiliary_resident_weight_gb": 11.197441024, "fixed_weight_gb": 38.84220416, "routed_expert_weight_gb": 12.33125376, "routed_experts": 48, "routed_experts_per_token": 3, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16", "traffic_scope": "ordinary text decode through model.layers.0-60, model.norm, and lm_head, with fused routed expert tensors divided into uniform expert-index groups and expected distinct routed experts across concurrent sessions", "auxiliary_scope": "model.embed_tokens.weight, vision_model, vit_downsampler, vit_downsampler2, and vit_large_projector tensors are resident for the multimodal package but not swept as full matrices for each ordinary cached text decode token", "shared_expert_notes": "MoE layers 4-59 instantiate one shared Step3vMLP with share_expert_dim 5120. Shared expert tensors are included in fixed_weight_gb because they are always read when a MoE layer runs.", "notes": "Layers 4-59 are MoE layers. The checkpoint stores routed experts as fused per-layer BF16 tensors named moe.down_proj, moe.gate_proj, and moe.up_proj with expert dimension 48, so routed_expert_weight_gb is the grouped routed tensor byte count divided by 48 expert indexes. Router gate tensors are included in fixed ordinary decode traffic." }, "kv_adapter": { "kind": "full_context", "layers": 61, "kv_heads": 1, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Step3vAttention hard-codes one key/value head, repeats it across 64 query heads, and every decoder layer is marked full_attention. The served config and deployment guide do not declare a quantized KV cache for the BF16 checkpoint, so Bounds Engine v1 charges BF16 full-context K/V cache traffic." }, "notes": "Step3vForConditionalGeneration wraps a StepRobotics vision encoder, downsamplers/projector, Step3Text MoE language model, and lm_head. This profile models ordinary cached text decode after any multimodal prefill; vision encoder and image-projector throughput are outside Bounds Engine v1." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-bf16-step3-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo. Activation traffic, attention kernels, router compute, expert compute, expert-parallel communication, cache writes, vision encoder throughput, and image prefill are outside this memory-side bound.", "notes": "The live API safetensors block and shard-header audit record only BF16 tensors. The model card says BF16 inference is currently supported, and the deployment guide documents BF16 vLLM/SGLang launch commands." }, "evidence": [ { "label": "Step3 API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/step3", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At commit e2d67fc936dc60f8f606f2429728f1dea503e10d, the live API records a public non-gated Apache-2.0 image-text-to-text repo with transformers, safetensors, step3_vl, custom_code, arxiv:2507.19427, endpoints_compatible, and region:us tags. Current downloads are 164103. The API safetensors block reports BF16 320969912832 parameters." }, { "label": "Step3 model card", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/README.md", "source_type": "model_card", "supports": [ "architecture", "serving" ], "notes": "The model card identifies Step3 as a VLM and says the current inference path supports BF16 inference, with vLLM and SGLang deployment guidance linked from the card." }, { "label": "Step3 deployment guide", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/docs/deploy_guidance.md", "source_type": "vendor_doc", "supports": [ "serving", "kv_adapter", "resident_weight_gb" ], "notes": "The deployment guide describes Step3 as a 321B-parameter VLM, says the BF16 version requires about 642 GB memory, documents vLLM/SGLang BF16 serving commands, and explicitly notes that Step3 has a single KV head." }, { "label": "Step3 config", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The config records Step3VLForConditionalGeneration, model_type step3_vl, BF16 text dtype, hidden size 7168, 61 text layers, 65536 max context, 64 attention heads, head_dim 256, share_q_dim 2048, MoE layers 4-59, 48 experts, top-k 3, moe_intermediate_size 5120, share_expert_dim 5120, vocabulary size 128815, and a 63-layer vision encoder." }, { "label": "Step3 remote modeling code", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/modeling_step3.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "routed_experts", "shared_experts_per_token", "auxiliary_resident_scope" ], "notes": "Manual review found Step3vAttention hard-codes num_key_value_heads = 1, projects K/V to one 256-wide head, repeats K/V across 64 query heads, and uses full_attention for every Step3vDecoderLayer. Step3vMoEMLP uses fused MoELinear tensors with shape [num_experts, out_features, in_features], top-k routing, and a separate shared expert in MoE layers. Step3vModel wraps vision_model, Step3Model, vit_downsampler, vit_downsampler2, and vit_large_projector." }, { "label": "Step3 safetensors index and shard headers", "url": "https://huggingface.co/stepfun-ai/step3/resolve/e2d67fc936dc60f8f606f2429728f1dea503e10d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split", "embedding_layout" ], "notes": "The index maps 1663 tensors across 89 shards and records total_size 641945818112. Direct range-read safetensors headers found BF16 tensor payloads totaling 641.939825664 GB, with 0.000204096 GB header/container overhead. Resident-only input embedding, vision encoder, downsamplers, and projector tensors sum to 11.197441024 GB. Main ordinary text resident bytes sum to 630.742384640 GB. Fused routed expert tensors sum to 591.900180480 GB, or 12.331253760 GB per expert index across 48 experts. Fixed ordinary text traffic, including dense layers, self-attention, router gates, shared experts, norms, and lm_head, sums to 38.842204160 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live Hugging Face API metadata, model card, deployment guide, pinned config, remote modeling code, safetensors index, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated dense estimate. It models the BF16 Step3 checkpoint as a sparse MoE VLM with exact fused expert byte groups and one-head BF16 full-context KV cache." }, { "id": "sugoitoolkit--sugoi-14b-ultra-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "sugoitoolkit/Sugoi-14B-Ultra-GGUF", "title": "Sugoi 14B Ultra GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Sugoi 14B Ultra.", "model_family": "qwen2-5-14b-dense-gguf", "base_model_proof": { "base_model": "sugoitoolkit/Sugoi-14B-Ultra-HF", "relation": "quantized", "source": "Hugging Face model card/API metadata, pinned Sugoi 14B Ultra HF config, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo metadata identifies sugoitoolkit/Sugoi-14B-Ultra-HF as the quantized base, and the base repo identifies Qwen/Qwen2.5-14B-Instruct as its finetuned source. The selected GGUF header matches the pinned HF config on Qwen2 architecture, 48 layers, 32768 context, 5120 hidden size, 13824 feed-forward size, 40 attention heads, 8 KV heads, and untied embeddings." }, "architecture": { "canonical_architecture_id": "qwen2-5-14b-sugoi-ultra", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 14.770033664, "swept_params_b": 13.991465984, "auxiliary_resident_params_b": 0.77856768, "resident_weight_gb": 29.54771584, "swept_weight_gb": 27.984613376, "auxiliary_resident_weight_gb": 1.563102464, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Sugoi-14B-Ultra-F16.gguf", "swept_parameter_scope": "ordinary text decode excludes token_embd.weight input lookup and includes blk.* tensors, output.weight, and output_norm.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file alignment are resident in the selected artifact but not swept as full matrices for each generated token", "notes": "The selected F16 linked file is 29.547715840 GB. Header tensor spans total 29.541748736 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005967104 GB. The GGUF stores token_embd.weight separately from output.weight, so ordinary text decode excludes the input embedding lookup and charges the separate output projection." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The pinned HF config records use_sliding_window false. The selected GGUF header records 48 Qwen2 blocks, 40 attention heads, and 8 KV heads. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. Smaller Q8_0, Q4_K_M, and Q2_K files in the repo require separate selected-artifact profiles." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.000517839848845, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains Q8_0, Q4_K_M, and Q2_K quantized GGUF siblings. This profile intentionally targets Sugoi-14B-Ultra-F16.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "Sugoi 14B Ultra GGUF API metadata", "url": "https://huggingface.co/api/models/sugoitoolkit/Sugoi-14B-Ultra-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit d8dd836bf519a604530779cbf5ce13ffb35c3eeb, the API records a public non-gated Apache-2.0 translation GGUF repo with base_model sugoitoolkit/Sugoi-14B-Ultra-HF, ja/en tags, region:us, 187132 downloads, GGUF architecture qwen2, 32768 context length, gguf.total 14770033664, and gguf.totalFileSize 29547715840." }, { "label": "Sugoi 14B Ultra GGUF model card", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-GGUF/raw/d8dd836bf519a604530779cbf5ce13ffb35c3eeb/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The card identifies the repo as the GGUF version of Sugoi 14B Ultra, records Apache-2.0 licensing, Japanese and English language tags, translation and GGUF tags, and base_model sugoitoolkit/Sugoi-14B-Ultra-HF. The card describes Japanese-to-English game-dialogue/localization use." }, { "label": "Sugoi 14B Ultra HF base config", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-HF/raw/9e71532aff2389b5e00f00eab72ee59f77de17a4/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "max_context_tokens", "embedding_layout" ], "notes": "The pinned base config records Qwen2ForCausalLM, qwen2, bfloat16 source weights, 48 layers, hidden size 5120, intermediate size 13824, 40 attention heads, 8 KV heads, use_sliding_window false, tie_word_embeddings false, vocab size 152064, max_position_embeddings 32768, and use_cache true." }, { "label": "Sugoi 14B Ultra GGUF linked-object HEAD checks", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-GGUF/tree/d8dd836bf519a604530779cbf5ce13ffb35c3eeb", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Sugoi-14B-Ultra-F16.gguf is 29547715840 bytes, exactly matching API gguf.totalFileSize. Sibling linked sizes are Q8_0 15701597440 bytes, Q4_K_M 8988110080 bytes, and Q2_K 5770497280 bytes." }, { "label": "Sugoi 14B Ultra F16 GGUF range-read tensor index", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-GGUF/resolve/d8dd836bf519a604530779cbf5ce13ffb35c3eeb/Sugoi-14B-Ultra-F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "embedding_layout" ], "notes": "An 8MB range-read of the selected GGUF v3 header found 24 metadata entries and 579 tensors. The linked file is 29.547715840 GB. Tensor spans sum to 29.541748736 GB: token_embd.weight 1.557135360 GB, blk.* tensors 26.427457536 GB, output.weight 1.557135360 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.005967104 GB. Tensor spans split into F16 29.538385920 GB and F32 0.003362816 GB. The header records qwen2.block_count 48, context_length 32768, embedding_length 5120, feed_forward_length 13824, attention.head_count 40, attention.head_count_kv 8, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the pinned model card, pinned Sugoi 14B Ultra HF config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Sugoi 14B Ultra F16 GGUF artifact. Do not silently substitute Q8_0, Q4_K_M, or Q2_K; those require separate profiles with their own selected artifact bytes." }, { "id": "sugoitoolkit--sugoi-32b-ultra-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "sugoitoolkit/Sugoi-32B-Ultra-GGUF", "title": "Sugoi 32B Ultra GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Sugoi 32B Ultra.", "model_family": "qwen2.5-32b-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen2.5-32B-Instruct", "relation": "quantized", "source": "Hugging Face model card/API metadata, Qwen2.5 32B Instruct config, existing audited Qwen profile, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo metadata identifies Qwen/Qwen2.5-32B-Instruct as the quantized base. The selected GGUF header records the same Qwen2 architecture, layer count, context length, attention head count, KV head count, and embedding geometry as the audited Qwen2.5 32B Instruct profile." }, "architecture": { "canonical_architecture_id": "qwen2-5-32b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 32.763876352, "swept_params_b": 31.986427904, "auxiliary_resident_params_b": 0.777448448, "resident_weight_gb": 65.53596928, "swept_weight_gb": 63.972855808, "auxiliary_resident_weight_gb": 1.563113472, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Sugoi-32B-Ultra-F16.gguf", "swept_parameter_scope": "ordinary text decode excludes token_embd.weight input lookup and includes blk.* tensors, output.weight, and output_norm.weight from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file alignment are resident in the selected artifact but not swept as full matrices for each generated token", "notes": "The selected F16 linked file is 65.535969280 GB. Header tensor spans total 65.529991168 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005978112 GB. The main GGUF stores token_embd.weight separately from output.weight, so ordinary text decode excludes the input embedding lookup and charges the separate output projection." }, "kv_adapter": { "kind": "full_context", "layers": 64, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Qwen2.5 32B Instruct config records use_sliding_window false. The selected GGUF header records 64 Qwen2 blocks, 40 attention heads, and 8 KV heads. The 128 head dimension is derived from hidden_size 5120 divided by 40 attention heads." }, "notes": "This profile models ordinary text decode for the API-selected F16 GGUF artifact. Smaller Q8_0, Q4_K_M, and Q2_K files in the repo require separate selected-artifact profiles." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2.000250781559292, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo contains Q8_0, Q4_K_M, and Q2_K quantized GGUF siblings. This profile intentionally targets Sugoi-32B-Ultra-F16.gguf because the HF API gguf.totalFileSize exactly matches that linked object." }, "evidence": [ { "label": "Sugoi 32B Ultra GGUF API metadata", "url": "https://huggingface.co/api/models/sugoitoolkit/Sugoi-32B-Ultra-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 3167b2d18c29a12f6eece6f6941e8eeb52d29e08, the API records a public non-gated Apache-2.0 translation GGUF repo with base_model Qwen/Qwen2.5-32B-Instruct, ja/en tags, region:us, 251014 downloads, GGUF architecture qwen2, 32768 context length, gguf.total 32763876352, and gguf.totalFileSize 65535969280." }, { "label": "Sugoi 32B Ultra GGUF model card", "url": "https://huggingface.co/sugoitoolkit/Sugoi-32B-Ultra-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline" ], "notes": "The card identifies the repo as the GGUF version of Sugoi 32B Ultra, records Apache-2.0 licensing, Japanese and English language tags, translation tag, and base_model Qwen/Qwen2.5-32B-Instruct." }, { "label": "Qwen2.5 32B Instruct audited profile and config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter", "max_context_tokens", "embedding_layout" ], "notes": "The audited Qwen2.5 32B Instruct profile records Qwen2ForCausalLM, bfloat16 source weights, 64 layers, hidden size 5120, 40 attention heads, 8 KV heads, sliding_window 131072 with use_sliding_window false, tie_word_embeddings false, 32768 max positions, stored input embedding and separate lm_head tensors, and BF16 safetensors total 32763876352 parameters." }, { "label": "Sugoi 32B Ultra GGUF linked-object HEAD checks", "url": "https://huggingface.co/sugoitoolkit/Sugoi-32B-Ultra-GGUF/tree/3167b2d18c29a12f6eece6f6941e8eeb52d29e08", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Sugoi-32B-Ultra-F16.gguf is 65535969280 bytes, exactly matching API gguf.totalFileSize. Sibling linked sizes are Q8_0 34820884480 bytes, Q4_K_M 19851335680 bytes, and Q2_K 12313098240 bytes." }, { "label": "Sugoi 32B Ultra F16 GGUF range-read tensor index", "url": "https://huggingface.co/sugoitoolkit/Sugoi-32B-Ultra-GGUF/resolve/3167b2d18c29a12f6eece6f6941e8eeb52d29e08/Sugoi-32B-Ultra-F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "embedding_layout" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 24 metadata entries and 771 tensors. The linked file is 65.535969280 GB. Tensor spans sum to 65.529991168 GB: token_embd.weight 1.557135360 GB, blk.* tensors 62.415699968 GB, output.weight 1.557135360 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.005978112 GB. Tensor spans split into F16 65.525514240 GB and F32 0.004476928 GB. The header records qwen2.block_count 64, context_length 32768, embedding_length 5120, feed_forward_length 27648, attention.head_count 40, attention.head_count_kv 8, and a separate output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, existing audited Qwen2.5 32B Instruct config/profile evidence, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Sugoi 32B Ultra F16 GGUF artifact. Do not silently substitute Q4_K_M or Q2_K; those require separate profiles with their own selected artifact bytes." }, { "id": "thebloke--mistral-7b-instruct-v0-2-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/Mistral-7B-Instruct-v0.2-AWQ", "title": "TheBloke Mistral 7B Instruct v0.2 AWQ", "summary": "Audited memory-side text-decode bounds profile for TheBloke's AWQ Int4 package of Mistral 7B Instruct v0.2.", "model_family": "mistral-7b-dense-awq", "base_model_proof": { "base_model": "mistralai/Mistral-7B-Instruct-v0.2", "relation": "quantized", "source": "Hugging Face API metadata, AWQ model card, served AWQ config, current base config comparison, quant_config, and direct safetensors header", "config_compatible": true, "notes": "The AWQ repo API and card identify mistralai/Mistral-7B-Instruct-v0.2 as the quantized base model. Manual comparison against the current base config found matching checked text geometry and context fields; the AWQ artifact adds AWQ quantization metadata and changes torch_dtype to float16." }, "architecture": { "canonical_architecture_id": "mistral-7b-v0-2", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.241732096, "swept_params_b": 7.110660096, "auxiliary_resident_params_b": 0.131072, "resident_weight_gb": 4.150796288, "swept_weight_gb": 3.888652288, "auxiliary_resident_weight_gb": 0.262144, "resident_parameter_scope": "logical AWQ model parameters represented by the safetensors qweight plus F16 model tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes layer tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from the safetensors header because the AWQ package mixes packed I32 qweight/qzeros tensors, F16 scales, and unquantized F16 embedding/head/norm tensors. Logical parameter counts follow the Hugging Face API convention: I32 qweight tensors are counted as unpacked 4-bit logical parameters, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records sliding_window null, so this profile charges full-context K and V streams for all 32 language layers." }, "notes": "Dense MistralForCausalLM profile using the served AWQ config and direct safetensors header grouping." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.573177277614662, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, F16 scales, and unquantized F16 tensors from the safetensors header. AWQ dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "TheBloke Mistral 7B Instruct v0.2 AWQ API metadata", "url": "https://huggingface.co/api/models/TheBloke/Mistral-7B-Instruct-v0.2-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit f970a2bb89d5c2f9d217dc337f39e24625d6462a, the API reports a public non-gated Transformers text-generation repo with base_model mistralai/Mistral-7B-Instruct-v0.2, Apache-2.0 license, 4-bit and awq tags, region:us, current downloads 275123, and safetensors logical parameters I32: 6979321856, F16: 262410240, total: 7241732096." }, { "label": "TheBloke Mistral 7B Instruct v0.2 AWQ served config", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ/raw/f970a2bb89d5c2f9d217dc337f39e24625d6462a/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records MistralForCausalLM with hidden_size 4096, intermediate_size 14336, 32 layers, 32 attention heads, 8 KV heads, max_position_embeddings 32768, sliding_window null, tie_word_embeddings false, vocab_size 32000, torch_dtype float16, rope_theta 1000000, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "TheBloke Mistral 7B Instruct v0.2 AWQ model card and quant_config", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "context_notes", "runtime_format" ], "notes": "The model card states the package contains AWQ model files for Mistral 7B Instruct v0.2, released as a 128g GEMM 4-bit model, with vLLM/TGI/AutoAWQ support. quant_config.json records zero_point true, q_group_size 128, w_bit 4, and version GEMM." }, { "label": "Mistral 7B Instruct v0.2 base config", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/raw/63a8b081895390a26e140280378bc85ec8bce07a/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The current base config records the same checked architecture fields as the AWQ config after excluding quantization_config, torch_dtype, and transformers_version: MistralForCausalLM, hidden_size 4096, intermediate_size 14336, 32 layers, 32 attention heads, 8 KV heads, max_position_embeddings 32768, sliding_window null, tie_word_embeddings false, vocab_size 32000, and rope_theta 1000000." }, { "label": "TheBloke Mistral 7B Instruct v0.2 AWQ safetensors header", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ/resolve/f970a2bb89d5c2f9d217dc337f39e24625d6462a/model.safetensors", "source_type": "manual_review", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "A direct range read found an 83936-byte safetensors header and 739 tensors totaling 4.150796288 GB: 3.516923904 GB I32 tensors and 0.633872384 GB F16 tensors. Stored suffix totals are qweight 3.489660928 GB, qzeros 0.027262976 GB, scales 0.109051904 GB, and F16 weight tensors 0.524820480 GB. model.embed_tokens.weight and lm_head.weight each have shape [32000, 4096] and contribute 0.262144000 GB. The linked file size is 4.150880232 GB, or 83944 bytes of safetensors header/container overhead above tensor payload." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served AWQ config, model card, quant_config, current base config comparison, linked-object HEAD check, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped flat 0.5 byte/parameter estimate by using exact stored AWQ tensor bytes and by excluding only the input embedding matrix from ordinary text-decode swept traffic." }, { "id": "thebloke--mistral-7b-instruct-v0-2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/Mistral-7B-Instruct-v0.2-GGUF", "title": "TheBloke Mistral 7B Instruct v0.2 GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-selected Q4_K_M GGUF artifact of Mistral 7B Instruct v0.2.", "model_family": "mistral-7b-dense-gguf", "base_model_proof": { "base_model": "mistralai/Mistral-7B-Instruct-v0.2", "relation": "quantized", "source": "Hugging Face model card/API metadata, selected Q4_K_M GGUF header metadata, stale repo-config check, pinned base config comparison, and linked-object size checks", "config_compatible": true, "notes": "The repo card and API metadata identify mistralai/Mistral-7B-Instruct-v0.2 as the quantized base. The repo-local config.json contains only model_type mistral, so this profile uses the selected Q4_K_M GGUF header plus the pinned base config for architecture evidence. Both record a Mistral/Llama-style dense decoder with 32 layers, hidden size 4096, intermediate size 14336, 32 attention heads, 8 KV heads, 128 key/value head dimension, 32768-token context, null sliding_window, and untied token/output embeddings." }, "architecture": { "canonical_architecture_id": "mistral-7b-v0-2", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.241732096, "swept_params_b": 7.110660096, "auxiliary_resident_params_b": 0.131072, "resident_weight_gb": 4.368439584, "swept_weight_gb": 4.293976064, "auxiliary_resident_weight_gb": 0.07446352, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for mistral-7b-instruct-v0.2.Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q4_K_M linked file is 4.368439584 GB. GGUF tensor spans total 4.367704064 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.000735520 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. The selected artifact mixes Q4_K, Q6_K, and F32 tensor classes, so exact tensor spans drive the bound instead of a flat 4-bit estimate." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The pinned base config records sliding_window null, and the selected GGUF header records 32 decoder layers, 8 KV heads, 128-dimensional key/value heads, and 32768 context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the card-selected Q4_K_M GGUF artifact. Other GGUF quantizations in this repo have different resident and traffic bytes and require separate workload selection." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.603231316222389, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, llama.cpp kernels, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The rendered model card's text-generation-webui download and llama.cpp run examples select mistral-7b-instruct-v0.2.Q4_K_M.gguf. The HF API gguf.totalFileSize points at the smaller Q2_K artifact, so that API field is recorded as mismatched and is not used as the selected artifact." }, "evidence": [ { "label": "TheBloke Mistral 7B Instruct v0.2 GGUF API metadata", "url": "https://huggingface.co/api/models/TheBloke/Mistral-7B-Instruct-v0.2-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "license", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "At commit 3a6fbf4a41a1d52e415a4958cde6856d34b2db93, the API records a public non-gated Apache-2.0 GGUF text-generation repo with base_model mistralai/Mistral-7B-Instruct-v0.2, base_model:quantized metadata, region:us, current downloads 86403, GGUF architecture llama, context_length 32768, gguf.total 7241732096, and gguf.totalFileSize 3083098400. The API totalFileSize matches mistral-7b-instruct-v0.2.Q2_K.gguf, while this profile targets Q4_K_M because the card's download and llama.cpp examples select that file." }, { "label": "TheBloke Mistral 7B Instruct v0.2 GGUF model card", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/raw/3a6fbf4a41a1d52e415a4958cde6856d34b2db93/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "selected_artifact" ], "notes": "The pinned card records Apache-2.0 licensing, model creator Mistral AI, original model mistralai/Mistral-7B-Instruct-v0.2, GGUF runtime guidance, and available 2/3/4/5/6/8-bit GGUF variants. The provided files table marks Q4_K_M as the balanced recommended medium option, and the download plus llama.cpp examples use mistral-7b-instruct-v0.2.Q4_K_M.gguf with 32768 context." }, { "label": "Mistral 7B Instruct v0.2 base config", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/raw/63a8b081895390a26e140280378bc85ec8bce07a/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "attention_heads", "kv_heads", "head_dim", "max_context_tokens", "tie_word_embeddings" ], "notes": "The pinned base config records MistralForCausalLM, bfloat16 source dtype, hidden size 4096, intermediate size 14336, 32 hidden layers, 32 attention heads, 8 KV heads, 32768 max position embeddings, sliding_window null, rope_theta 1000000, vocab_size 32000, and tie_word_embeddings false." }, { "label": "TheBloke Mistral 7B Instruct v0.2 GGUF linked-object HEAD checks", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/tree/3a6fbf4a41a1d52e415a4958cde6856d34b2db93", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 3.083098400 GB, Q3_K_S 3.164567840 GB, Q3_K_M 3.518986528 GB, Q3_K_L 3.822024992 GB, Q4_0 4.108917024 GB, Q4_K_S 4.140374304 GB, Q4_K_M 4.368439584 GB, Q5_0 4.997716256 GB, Q5_K_S 4.997716256 GB, Q5_K_M 5.131409696 GB, Q6_K 5.942065440 GB, and Q8_0 7.695857952 GB. The selected Q4_K_M artifact intentionally differs from the API gguf.totalFileSize artifact because the rendered examples choose Q4_K_M." }, { "label": "TheBloke Mistral 7B Instruct v0.2 Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/resolve/3a6fbf4a41a1d52e415a4958cde6856d34b2db93/mistral-7b-instruct-v0.2.Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "attention_heads", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 24 metadata entries and 291 tensors. The linked file is 4.368439584 GB. Tensor spans sum to 4.367704064 GB across 7.241732096B logical elements: output.weight 0.107520000 GB, output_norm.weight 0.000016384 GB, token_embd.weight 0.073728000 GB, and blk.* tensors 4.186439680 GB. Metadata/tokenizer/header/file overhead accounts for 0.000735520 GB. Stored tensor spans split into Q4_K 3.433365504 GB, Q6_K 0.933273600 GB, and F32 0.001064960 GB. The header records general.architecture llama, llama.block_count 32, context_length 32768, embedding_length 4096, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, rope.dimension_count 128, rope.freq_base 1000000, and a separate output.weight tensor." }, { "label": "TheBloke Mistral 7B Instruct v0.2 GGUF stale config", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/raw/3a6fbf4a41a1d52e415a4958cde6856d34b2db93/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The repo-local config.json contains only model_type mistral and omits layer count, attention geometry, context length, and embedding layout. It is recorded as stale/incomplete and is not used as architecture evidence." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned base config, linked GGUF file sizes, stale repo-config check, and direct selected Q4_K_M GGUF tensor-index range read." }, "notes": "Use this profile for the card-selected Mistral 7B Instruct v0.2 Q4_K_M GGUF artifact in ordinary text-decode bounds. Do not infer Q2_K or other quantized sibling footprints from this profile." }, { "id": "thebloke--tinyllama-1-1b-chat-v0-3-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", "title": "TheBloke TinyLlama 1.1B Chat v0.3 AWQ", "summary": "Audited memory-side text-decode bounds profile for TheBloke's AWQ Int4 package of TinyLlama 1.1B Chat v0.3.", "model_family": "tinyllama-llama-dense-awq", "base_model_proof": { "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.3", "relation": "quantized", "source": "Hugging Face API metadata, AWQ model card, served AWQ config, current base config comparison, quant_config, and direct safetensors header", "config_compatible": true, "notes": "The old model card metadata names PY007/TinyLlama-1.1B-Chat-v0.3 as the original model, while the current Hub API resolves both PY007/TinyLlama-1.1B-Chat-v0.3 and the repo tags to TinyLlama/TinyLlama-1.1B-Chat-v0.3. Manual comparison against the current base config found matching checked text geometry and context fields; the AWQ artifact adds AWQ quantization metadata and changes torch_dtype to float16." }, "architecture": { "canonical_architecture_id": "tinyllama-1-1b-chat-v0-3", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.100060672, "swept_params_b": 1.034518528, "auxiliary_resident_params_b": 0.065542144, "resident_weight_gb": 0.765718528, "swept_weight_gb": 0.63463424, "auxiliary_resident_weight_gb": 0.131084288, "resident_parameter_scope": "logical AWQ model parameters represented by qweight tensors plus F16 model tensors in model.safetensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers tensors, model.norm.weight, and lm_head.weight plus AWQ qzeros/scales storage traffic", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full input embedding matrix for each ordinary decode token", "notes": "AWQ qweight tensors are packed I32 values; logical parameter counts treat each qweight element as eight 4-bit values. qzeros and scales are storage/runtime overhead and are charged in stored-byte traffic but excluded from logical model parameter counts." }, "kv_adapter": { "kind": "full_context", "layers": 22, "kv_heads": 4, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The served AWQ config records 22 decoder layers, 4 KV heads, hidden size 2048, and 32 attention heads, so head_dim is 64. No sliding-window or recurrent state is recorded." }, "notes": "Dense LlamaForCausalLM text-generation profile for the main AWQ branch. It models ordinary autoregressive text decode with no speculative decoding." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6960693600725325, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, qzeros, F16 scales, and unquantized F16 tensors from the safetensors header. AWQ dequantization, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "TheBloke TinyLlama 1.1B Chat v0.3 AWQ API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "pipeline", "license", "downloads", "safetensors_logical_parameter_split", "commit_sha" ], "notes": "At commit d87af942d412211c2fe4a6cd6c4964cdd7e93e22, the API reports a public non-gated Apache-2.0 Transformers text-generation repo with base_model TinyLlama/TinyLlama-1.1B-Chat-v0.3, 4-bit and awq tags, text-generation-inference, deploy:azure, region:us, current downloads 132292, and safetensors logical parameters I32: 968884224, F16: 131176448, total: 1100060672." }, { "label": "TheBloke TinyLlama 1.1B Chat v0.3 AWQ served config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ/raw/d87af942d412211c2fe4a6cd6c4964cdd7e93e22/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization" ], "notes": "The config records LlamaForCausalLM with hidden_size 2048, intermediate_size 5632, 22 layers, 32 attention heads, 4 KV heads, max_position_embeddings 2048, rope_theta 10000, tie_word_embeddings false, vocab_size 32003, torch_dtype float16, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "TheBloke TinyLlama 1.1B Chat v0.3 AWQ model card and quant_config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", "source_type": "model_card", "supports": [ "base_model_proof", "quantization", "context_notes", "runtime_format" ], "notes": "The card states the package contains AWQ model files for TinyLlama 1.1B Chat v0.3, released as a sharded safetensors 128g 4-bit AWQ model with 2048 sequence length and vLLM/TGI/AutoAWQ support. quant_config.json records zero_point true, q_group_size 128, w_bit 4, and version GEMM." }, { "label": "TinyLlama 1.1B Chat v0.3 base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3/raw/1ea93498c83d6d5aece90cf3b160aadf9f8875f8/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture_compatibility", "kv_adapter" ], "notes": "The current base config records the same checked architecture fields as the AWQ config after excluding quantization_config, torch_dtype, and transformers_version: LlamaForCausalLM, hidden_size 2048, intermediate_size 5632, 22 layers, 32 attention heads, 4 KV heads, max_position_embeddings 2048, tie_word_embeddings false, vocab_size 32003, and rope_theta 10000. The base config records torch_dtype float32." }, { "label": "TheBloke TinyLlama 1.1B Chat v0.3 AWQ safetensors header", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ/resolve/d87af942d412211c2fe4a6cd6c4964cdd7e93e22/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "Direct safetensors header inspection found a 56456-byte header with 509 tensors and 0.765718528 GB of tensor payload. The linked file size is 0.765774992 GB, or 56464 bytes of safetensors header/container overhead above tensor payload. Dtype bytes are I32 0.488226816 GB and F16 0.277491712 GB. Stored suffix bytes are qweight 0.484442112 GB, qzeros 0.003784704 GB, scales 0.015138816 GB, and F16 weight tensors 0.262352896 GB. model.embed_tokens.weight and lm_head.weight each have shape [32003, 2048] and contribute 0.131084288 GB / 65.542144M logical parameters. model.layers tensors contribute 0.503545856 GB, and model.norm.weight contributes 0.000004096 GB. Ordinary text swept traffic is layers plus norm plus lm_head, totaling 0.634634240 GB and 1.034518528B logical serving parameters." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, served AWQ config, quant_config, current TinyLlama base config comparison, and direct safetensors header byte grouping." }, "notes": "This profile supersedes the scraped flat 0.5 byte/parameter estimate by using exact stored AWQ tensor bytes and by excluding only the input embedding matrix from ordinary text-decode swept traffic." }, { "id": "thebloke--tinyllama-1-1b-chat-v0-3-gptq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", "title": "TheBloke TinyLlama 1.1B Chat v0.3 GPTQ", "summary": "Audited memory-side text-decode bounds profile for the main-branch GPTQ 4-bit TinyLlama 1.1B Chat v0.3 artifact.", "model_family": "tinyllama-llama-dense", "base_model_proof": { "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.3", "relation": "quantized", "source": "Hugging Face model metadata, model card, served config comparison, and quantize_config.json", "config_compatible": true, "notes": "The model card records PY007/TinyLlama-1.1B-Chat-v0.3 as the original model, while the current Hub API resolves the base repo as TinyLlama/TinyLlama-1.1B-Chat-v0.3. Manual comparison found matching checked architecture fields between the GPTQ config and the base config; the GPTQ repo changes dtype labels and adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "tinyllama-1-1b-chat-v0-3", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.100454912, "swept_params_b": 1.034912768, "auxiliary_resident_params_b": 0.065542144, "resident_weight_gb": 0.768083968, "swept_weight_gb": 0.63699968, "auxiliary_resident_weight_gb": 0.131084288, "resident_parameter_scope": "GPTQ logical serving parameters reconstructed from qweight tensors plus non-scale F16 tensors in model.safetensors", "swept_parameter_scope": "model.layers tensors, model.norm.weight, and lm_head.weight plus GPTQ per-layer metadata stored in the safetensors header", "auxiliary_scope": "model.embed_tokens.weight is resident for ordinary text decode but not swept as a full input embedding matrix for each generated token", "notes": "GPTQ qweight tensors are packed I32 values; logical parameter counts treat each qweight element as eight 4-bit values. qzeros, scales, and g_idx are storage/serving metadata and are charged in stored-byte traffic but excluded from logical parameter counts. AutoGPTQ stores F16 bias tensors for quantized modules; these are included as swept serving traffic and logical serving parameters." }, "kv_adapter": { "kind": "full_context", "layers": 22, "kv_heads": 4, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 22 decoder layers, 4 KV heads, hidden size 2048, and 32 attention heads, so head_dim is 64. No sliding-window or recurrent state is recorded." }, "notes": "This is a LlamaForCausalLM text-generation profile for the main GPTQ branch. It models ordinary autoregressive text decode with no speculative decoding." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.6979695029976839, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "gptq-4bit-group128-actorder-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ bytes: packed I32 qweight/qzeros/g_idx tensors plus F16 scales, embeddings, output head, norms, and bias tensors. Dequantization, activation traffic, compute, and kernel scheduling overhead are outside Bounds Engine v1.", "notes": "The main branch quantize_config records 4-bit GPTQ, group_size 128, desc_act true, damp_percent 0.1, symmetric quantization, and true_sequential true. KV cache is charged at FP16." }, "evidence": [ { "label": "TheBloke TinyLlama GPTQ model card and API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 05835c52707fff57cecd16a364de2bc65c9bf102, the API records a public Apache-2.0 text-generation repo with transformers, safetensors, llama, text-generation-inference, 4-bit, gptq, deploy:azure, and region:us tags. Current downloads are 1068629. The API safetensors block reports I32 968884224 logical quantized parameters and F16 131570688 non-scale parameters, total 1100454912." }, { "label": "TheBloke TinyLlama GPTQ config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/raw/05835c52707fff57cecd16a364de2bc65c9bf102/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, model_type llama, torch_dtype float16, hidden_size 2048, intermediate_size 5632, 22 hidden layers, 32 attention heads, 4 KV heads, 2048 max position embeddings, rope_theta 10000, vocab_size 32003, tie_word_embeddings false, and GPTQ quantization_config with bits 4, group_size 128, desc_act true, symmetric quantization, and true_sequential true." }, { "label": "TheBloke TinyLlama GPTQ quantize_config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/raw/05835c52707fff57cecd16a364de2bc65c9bf102/quantize_config.json", "source_type": "config", "supports": [ "serving", "weight_format" ], "notes": "quantize_config.json independently records the main-branch GPTQ settings: 4 bits, group_size 128, damp_percent 0.1, desc_act true, sym true, true_sequential true, and model_file_base_name model." }, { "label": "TinyLlama 1.1B Chat v0.3 base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3/raw/1ea93498c83d6d5aece90cf3b160aadf9f8875f8/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in checked architecture fields between the base config and the GPTQ config: LlamaForCausalLM, llama, hidden size 2048, intermediate size 5632, 22 layers, 32 attention heads, 4 KV heads, 2048 positions, rope_theta 10000, vocab_size 32003, and untied embeddings. The base config records torch_dtype float32; the GPTQ config records float16 and quantization metadata." }, { "label": "TheBloke TinyLlama GPTQ safetensors header", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/resolve/05835c52707fff57cecd16a364de2bc65c9bf102/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "Range-reading model.safetensors found a 89256-byte header with 817 tensors and 0.768083968 GB of tensor payload. Dtype bytes are I32 0.489803776 GB and F16 0.278280192 GB. Stored suffix bytes are qweight 0.484442112 GB, F16 weight 0.262352896 GB, F16 scales 0.015138816 GB, qzeros 0.003784704 GB, g_idx 0.00157696 GB, and F16 bias 0.00078848 GB. model.embed_tokens.weight contributes 0.131084288 GB and 65.542144M resident-only parameters. model.layers plus model.norm contribute 0.505915392 GB and 969.370624M logical serving parameters. lm_head.weight contributes 0.131084288 GB and 65.542144M swept parameters. Ordinary text swept traffic is layers/norm plus lm_head, totaling 0.63699968 GB and 1.034912768B logical serving parameters." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, pinned GPTQ config, quantize_config, current base config comparison, and direct range-read safetensors header grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a flat ideal 4-bit dense checkpoint and undercounted GPTQ metadata, F16 embeddings/output head/norms/biases, and exact swept decode traffic." }, { "id": "thebloke--tinyllama-1-1b-chat-v1-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", "title": "TheBloke TinyLlama 1.1B Chat v1.0 GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-selected Q4_K_M GGUF artifact of TinyLlama 1.1B Chat v1.0.", "model_family": "tinyllama-llama-dense-gguf", "base_model_proof": { "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "relation": "quantized", "source": "Hugging Face model card/API metadata, pinned TinyLlama v1.0 base config, selected GGUF header metadata, and linked-object size checks", "config_compatible": true, "notes": "The repo card and API metadata identify TinyLlama/TinyLlama-1.1B-Chat-v1.0 as the quantized base. The repo-local config.json contains only model_type tinyllama, so architecture comes from the selected Q4_K_M GGUF header and the pinned base config. Both record a Llama-style dense decoder with 22 layers, hidden size 2048, 32 attention heads, 4 KV heads, 64 key/value head dimension, 5632 intermediate size, 2048-token context, and untied token/output embeddings." }, "architecture": { "canonical_architecture_id": "tinyllama-1-1b-chat-v1-0", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.100048384, "swept_params_b": 1.034512384, "auxiliary_resident_params_b": 0.065536, "resident_weight_gb": 0.668788096, "swept_weight_gb": 0.630214656, "auxiliary_resident_weight_gb": 0.03857344, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q4_K_M GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q4_K_M linked file is 0.668788096 GB. GGUF tensor spans total 0.667078656 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.001709440 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 22, "kv_heads": 4, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The base config and selected GGUF header record 22 decoder layers, 4 KV heads, 64-dimensional key/value heads, and 2048-token context. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the card-selected Q4_K_M GGUF artifact. Other GGUF quantizations in this repo have different resident and traffic bytes and require separate workload selection." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6079624366776943, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, llama.cpp kernels, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The rendered model card's text-generation-webui download and llama.cpp run examples select tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf. The HF API gguf.totalFileSize points at the smaller Q2_K artifact, so that API field is recorded as mismatched and is not used as the selected artifact." }, "evidence": [ { "label": "TheBloke TinyLlama 1.1B Chat v1.0 GGUF API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "license", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "At commit 52e7645ba7c309695bec7ac98f4f005b139cf465, the API records a public non-gated Apache-2.0 GGUF repo with base_model TinyLlama/TinyLlama-1.1B-Chat-v1.0, base_model:quantized metadata, region:us, GGUF architecture llama, context_length 2048, gguf.total 1100048384, gguf.totalFileSize 483116416, and current downloads 194341. The API totalFileSize matches tinyllama-1.1b-chat-v1.0.Q2_K.gguf, while this profile targets Q4_K_M because the card's download and llama.cpp examples select that file." }, { "label": "TheBloke TinyLlama 1.1B Chat v1.0 GGUF model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/raw/52e7645ba7c309695bec7ac98f4f005b139cf465/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format", "selected_artifact" ], "notes": "The pinned card records Apache-2.0 licensing, model creator TinyLlama, original model TinyLlama/TinyLlama-1.1B-Chat-v1.0, GGUF runtime guidance, and available 2/3/4/5/6/8-bit GGUF variants. The provided files table marks Q4_K_M as the balanced recommended medium option, and the download plus llama.cpp examples use tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf." }, { "label": "TinyLlama 1.1B Chat v1.0 base config and API metadata", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0/raw/fe8a4ea1ffedaf415f4da2f062534de366a451e6/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "max_context_tokens", "tie_word_embeddings" ], "notes": "The pinned base config records LlamaForCausalLM, model_type llama, BF16 source dtype, hidden size 2048, intermediate size 5632, 22 hidden layers, 32 attention heads, 4 KV heads, 2048 max position embeddings, rope_theta 10000, vocab_size 32000, and tie_word_embeddings false. The base API at the same audit time records public non-gated Apache-2.0 text-generation metadata and BF16 safetensors total 1100048384 parameters." }, { "label": "TheBloke TinyLlama 1.1B Chat v1.0 GGUF linked-object HEAD checks", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/tree/52e7645ba7c309695bec7ac98f4f005b139cf465", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Q2_K 0.483116416 GB, Q3_K_S 0.500315520 GB, Q3_K_M 0.550819200 GB, Q3_K_L 0.592500096 GB, Q4_0 0.637699456 GB, Q4_K_S 0.643728768 GB, Q4_K_M 0.668788096 GB, Q5_0 0.767001984 GB, Q5_K_S 0.767001984 GB, Q5_K_M 0.783017344 GB, Q6_K 0.904385920 GB, and Q8_0 1.170781568 GB. The selected Q4_K_M artifact intentionally differs from the API gguf.totalFileSize artifact because the rendered examples choose Q4_K_M." }, { "label": "TheBloke TinyLlama 1.1B Chat v1.0 Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/52e7645ba7c309695bec7ac98f4f005b139cf465/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 23 metadata entries and 201 tensors. The linked file is 0.668788096 GB. Tensor spans sum to 0.667078656 GB across 1.100048384B logical elements: output.weight 0.053760000 GB, output_norm.weight 0.000008192 GB, token_embd.weight 0.036864000 GB, and blk.* tensors 0.576446464 GB. Metadata/tokenizer/header/file overhead accounts for 0.001709440 GB. Stored tensor bytes split into Q4_K 0.514031616 GB, Q6_K 0.152678400 GB, and F32 0.000368640 GB. The header records general.architecture llama, llama.block_count 22, context_length 2048, embedding_length 2048, feed_forward_length 5632, attention.head_count 32, attention.head_count_kv 4, rope.dimension_count 64, and rope.freq_base 10000." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned TinyLlama base config, linked-object HEAD checks for all GGUF siblings, and a direct selected Q4_K_M GGUF tensor-index range read." }, "notes": "Use this profile for the card-selected TinyLlama 1.1B Chat v1.0 Q4_K_M GGUF artifact in ordinary text-decode bounds. Do not infer Q2_K or other sibling quantization footprints from this profile." }, { "id": "thebloke--tinyllama-1-1b-chat-v1-0-gptq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "title": "TheBloke TinyLlama 1.1B Chat v1.0 GPTQ", "summary": "Audited memory-side text-decode bounds profile for the main-branch GPTQ 4-bit TinyLlama 1.1B Chat v1.0 artifact.", "model_family": "tinyllama-llama-dense", "base_model_proof": { "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "relation": "quantized", "source": "Hugging Face model metadata, pinned model card, served config comparison, and quantize_config.json", "config_compatible": true, "notes": "The model card and API metadata record TinyLlama/TinyLlama-1.1B-Chat-v1.0 as the original/base model. Manual comparison found matching checked architecture fields between the GPTQ config and the base config; the GPTQ repo changes dtype labels and adds GPTQ quantization metadata." }, "architecture": { "canonical_architecture_id": "tinyllama-1-1b-chat-v1-0", "max_context_tokens": 2048, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.100442624, "swept_params_b": 1.034906624, "auxiliary_resident_params_b": 0.065536, "resident_weight_gb": 0.768059392, "swept_weight_gb": 0.636987392, "auxiliary_resident_weight_gb": 0.131072, "resident_parameter_scope": "GPTQ logical serving parameters reconstructed from qweight tensors plus non-scale BF16/F16 tensors in model.safetensors", "swept_parameter_scope": "model.layers tensors, model.norm.weight, and lm_head.weight plus GPTQ per-layer metadata stored in the safetensors header", "auxiliary_scope": "model.embed_tokens.weight is resident for ordinary text decode but not swept as a full input embedding matrix for each generated token", "notes": "GPTQ qweight tensors are packed I32 values; logical parameter counts treat each qweight element as eight 4-bit values. qzeros, scales, and g_idx are storage/serving metadata and are charged in stored-byte traffic but excluded from logical parameter counts. AutoGPTQ stores F16 bias tensors for quantized modules; these are included as swept serving traffic and logical serving parameters." }, "kv_adapter": { "kind": "full_context", "layers": 22, "kv_heads": 4, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The config records 22 decoder layers, 4 KV heads, hidden size 2048, and 32 attention heads, so head_dim is 64. No sliding-window or recurrent state is recorded." }, "notes": "This is a LlamaForCausalLM text-generation profile for the main GPTQ branch. It models ordinary autoregressive text decode with no speculative decoding." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.697954963983656, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "gptq-4bit-group128-actorder-text-decode-memory-bound", "dequantization_notes": "The memory-side bound charges stored GPTQ bytes: packed I32 qweight/qzeros/g_idx tensors plus BF16 weights, F16 scales/bias tensors, embeddings, output head, and norms. Dequantization, activation traffic, compute, and kernel scheduling overhead are outside Bounds Engine v1.", "notes": "The main branch quantize_config records 4-bit GPTQ, group_size 128, desc_act true, damp_percent 0.1, symmetric quantization, and true_sequential true. KV cache is charged at FP16." }, "evidence": [ { "label": "TheBloke TinyLlama v1.0 GPTQ API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "serving" ], "notes": "At commit 9d4580af0f21bccafd762dcc50d0c7bac6273584, the API records a public non-gated Apache-2.0 text-generation repo with transformers, safetensors, llama, text-generation-inference, 4-bit, gptq, and region:us tags. Current downloads are 132526. The API safetensors block reports I32 968884224, BF16 131164160, F16 394240, and total 1100442624." }, { "label": "TheBloke TinyLlama v1.0 GPTQ model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/raw/9d4580af0f21bccafd762dcc50d0c7bac6273584/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "serving", "weight_format" ], "notes": "The pinned card identifies TinyLlama/TinyLlama-1.1B-Chat-v1.0 as the original model and describes the repo as GPTQ files for GPU inference. Its provided-files table records the main branch as 4-bit, group size 128, act-order yes, damp 0.1, sequence length 2048, and size 0.77 GB." }, { "label": "TheBloke TinyLlama v1.0 GPTQ config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/raw/9d4580af0f21bccafd762dcc50d0c7bac6273584/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records LlamaForCausalLM, model_type llama, torch_dtype bfloat16, hidden_size 2048, intermediate_size 5632, 22 hidden layers, 32 attention heads, 4 KV heads, 2048 max position embeddings, rope_theta 10000, vocab_size 32000, tie_word_embeddings false, and GPTQ quantization_config with bits 4, group_size 128, desc_act true, symmetric quantization, and true_sequential true." }, { "label": "TheBloke TinyLlama v1.0 GPTQ quantize_config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/raw/9d4580af0f21bccafd762dcc50d0c7bac6273584/quantize_config.json", "source_type": "config", "supports": [ "serving", "weight_format" ], "notes": "quantize_config.json independently records the main-branch GPTQ settings: 4 bits, group_size 128, damp_percent 0.1, desc_act true, sym true, true_sequential true, and model_file_base_name model." }, { "label": "TinyLlama 1.1B Chat v1.0 base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0/raw/fe8a4ea1ffedaf415f4da2f062534de366a451e6/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in checked architecture fields between the base config and the GPTQ config: LlamaForCausalLM, llama, hidden size 2048, intermediate size 5632, 22 layers, 32 attention heads, 4 KV heads, 2048 positions, rope_theta 10000, vocab_size 32000, and untied embeddings. The base config records torch_dtype bfloat16; the GPTQ config records the same dtype plus quantization metadata." }, { "label": "TheBloke TinyLlama v1.0 GPTQ safetensors header", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/resolve/9d4580af0f21bccafd762dcc50d0c7bac6273584/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "dtype_split", "storage_overhead" ], "notes": "Range-reading model.safetensors found an 89304-byte header with 817 tensors and 0.768059392 GB of tensor payload. Dtype bytes are I32 0.489803776 GB, BF16 0.262328320 GB, and F16 0.015927296 GB. Stored suffix bytes are qweight 0.484442112 GB, BF16 weights 0.262328320 GB, F16 scales 0.015138816 GB, qzeros 0.003784704 GB, g_idx 0.001576960 GB, and F16 bias 0.000788480 GB. model.embed_tokens.weight contributes 0.131072000 GB and 65.536000M resident-only parameters. model.layers plus model.norm contribute 0.505915392 GB and 969.370624M logical serving parameters. lm_head.weight contributes 0.131072000 GB and 65.536000M swept parameters. Ordinary text swept traffic is layers/norm plus lm_head, totaling 0.636987392 GB and 1.034906624B logical serving parameters. The linked file is 0.768148704 GB, leaving 0.000089312 GB of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned GPTQ config, quantize_config, pinned base config comparison, and direct range-read safetensors header grouping." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the repo as a flat ideal 4-bit dense checkpoint and undercounted GPTQ metadata, BF16 embeddings/output head/norms, F16 scales/biases, and exact swept decode traffic." }, { "id": "txn545--qwen3-5-122b-a10b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "txn545/Qwen3.5-122B-A10B-NVFP4", "title": "txn545 Qwen3.5 122B A10B NVFP4", "summary": "Audited memory-side text-decode bounds profile for txn545's ModelOpt NVFP4 Qwen3.5 122B A10B artifact.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-122B-A10B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config, hf_quant_config, audited BF16 base config comparison, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo records Qwen/Qwen3.5-122B-A10B as its base model and preserves the audited Qwen3.5 122B A10B text geometry: Qwen3_5MoeForConditionalGeneration, 48 text layers, 12 full-attention layers, 36 DeltaNet linear-attention layers, 2 KV heads, 256 full-attention head dimension, 256 experts, 8 routed experts per token, one shared expert, 262144 max positions, and the same DeltaNet state geometry. The quantized repo adds ModelOpt NVFP4 and FP8-KV metadata." }, "architecture": { "canonical_architecture_id": "qwen3-5-122b-a10b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 82.823706304, "main_resident_weight_gb": 75.348088032, "auxiliary_resident_weight_gb": 7.475618272, "fixed_weight_gb": 10.117977312, "routed_expert_weight_gb": 0.25480512, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for each ordinary text decode token", "shared_expert_notes": "The text config records shared_expert_intermediate_size 1024. The ModelOpt ignore list leaves shared_expert_gate unquantized, and shared expert traffic is included in fixed_weight_gb because it is always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the package mixes packed U8 NVFP4 tensors, F8_E4M3 scale tensors, BF16 tensors, and tiny F32 scale tensors. Routed experts are byte-uniform across 256 expert indexes; routed_expert_weight_gb is the grouped ordinary-language routed tensor byte count divided by 256." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 48 layers with every fourth layer using full_attention, giving 12 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154533888, "read_gb_per_output_token": 0.154533888, "state_formula": "36 linear_attention layers * ((64 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 64 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as BF16. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput and MTP speculative decode are outside this text-decode profile." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with resident-only multimodal, MTP, and input-embedding tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-sglang-modelopt-nvfp4-fp8-kv-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored NVFP4/FP8/BF16/F32 safetensors bytes and FP8 KV bytes. NVFP4 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, speculative MTP decode, and state writes are outside this memory-side bound.", "notes": "The served config records a ModelOpt quantization config with an 8-bit float KV cache scheme, and hf_quant_config records quant_algo NVFP4 with kv_cache_quant_algo FP8. weight_bytes_per_param records the nominal NVFP4 weight payload; the audited adapter uses exact stored tensor bytes for resident and per-token weight traffic." }, "evidence": [ { "label": "txn545 Qwen3.5 122B A10B NVFP4 API metadata and model card", "url": "https://huggingface.co/api/models/txn545/Qwen3.5-122B-A10B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "serving", "total_params_b", "active_params_b", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "At commit cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9, the API records a public Apache-2.0 text-generation Model Optimizer artifact derived from Qwen/Qwen3.5-122B-A10B, with current downloads 129110, region:us, and safetensors parameters BF16 8669395184, F8_E4M3 7276068864, U8 58208550912, total 64354266864. The tags identify ModelOpt, Qwen3.5, FP4/fp4, and base_model Qwen/Qwen3.5-122B-A10B." }, { "label": "txn545 Qwen3.5 122B A10B NVFP4 config", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/raw/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "auxiliary_resident_scope", "kv_store_format", "kv_read_format", "serving" ], "notes": "The config records Qwen3_5MoeForConditionalGeneration with qwen3_5_moe_text, 48 text layers, layer_types with every fourth layer full_attention, 12 full-attention layers, 36 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 64 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 1024, 262144 max position embeddings, a resident vision config, top-level MTP tensors, and a ModelOpt quantization_config with kv_cache_scheme {dynamic:false, num_bits:8, type:'float'}." }, { "label": "txn545 Qwen3.5 122B A10B NVFP4 hf_quant_config", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/raw/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "ignored_quantized_modules" ], "notes": "The ModelOpt sidecar records quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and excludes lm_head, shared expert gates, linear-attention modules, self-attention modules on full-attention layers, model.visual, and mtp.layers.0 from NVFP4 weight quantization." }, { "label": "Qwen3.5 122B A10B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found matching audited top-level, text, vision, MoE, attention, and DeltaNet state geometry fields between the txn545 NVFP4 config and the BF16 base config. The base API records Apache-2.0 metadata, current downloads 784041 when audited, region:us, and safetensors parameters BF16 125086490096 plus F32 6912." }, { "label": "txn545 Qwen3.5 122B A10B NVFP4 safetensors headers", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/resolve/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across both indexed shards. Stored tensors sum to the index total_size, 82.823706304 GB, across 149765 tensors: 58.208550912 GB U8, 7.276068864 GB F8_E4M3, 17.338790368 GB BF16, and 0.000296160 GB F32. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head.weight, sum to 75.348088032 GB. Auxiliary resident tensors, defined as model.visual plus top-level mtp plus model.language_model.embed_tokens.weight, sum to 7.475618272 GB. Ordinary-language routed expert tensors sum to 65.230110720 GB and divide exactly into 256 uniform expert indexes of 0.254805120 GB. Fixed ordinary text traffic sums to 10.117977312 GB. All index weight_map entries were found in the shard headers." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, reads cached conv_states and recurrent_states during decode, and updates recurrent state after the gated-delta rule." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served config, hf_quant_config, BF16 base config comparison, direct safetensors header byte grouping, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile is for ordinary text decode bounds. It deliberately separates resident visual/MTP/input-embedding weights from per-token swept language/logit weights and includes a fixed-state charge for DeltaNet linear-attention layers. The full resident package fits on some 128GB local hardware, but concurrency is constrained by FP8 full-attention KV plus fixed DeltaNet state." }, { "id": "unsloth--diffusiongemma-26b-a4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "unsupported", "repo": "unsloth/diffusiongemma-26B-A4B-it-GGUF", "title": "Unsloth DiffusionGemma 26B A4B IT GGUF BF16", "summary": "Unsupported profile stub with exact resident tensor evidence for the Unsloth BF16 GGUF DiffusionGemma package.", "model_family": "diffusion-gemma-block-diffusion-moe", "base_model_proof": { "base_model": "google/diffusiongemma-26B-A4B-it", "relation": "quantized", "source": "Hugging Face cardData base_model metadata, model card, API GGUF metadata, GGUF header metadata, and existing Google DiffusionGemma base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/diffusiongemma-26B-A4B-it. The selected GGUF header records the DiffusionGemma block-diffusion architecture, 30 decoder blocks, 262144-token context, 128 experts, 8 active experts, non-causal/bidirectional generation-canvas attention, and hybrid local/global attention metadata." }, "architecture": { "canonical_architecture_id": "diffusion-gemma-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 50.5408344, "main_resident_weight_gb": 50.525009792, "auxiliary_resident_weight_gb": 0.015824608, "fixed_weight_gb": 4.849023872, "routed_expert_weight_gb": 0.35684364, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected BF16 GGUF linked file size for diffusiongemma-26B-A4B-it-BF16.gguf", "traffic_scope": "Exact GGUF tensor byte groups are recorded here, but Bounds Engine v1 does not use them for production throughput because DiffusionGemma block diffusion is not ordinary one-output-token autoregressive decode.", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact but not swept as model tensors.", "shared_expert_notes": "The GGUF metadata records 128 routed experts and 8 active experts. Dense blk.* non-expert tensors plus self-conditioning, embeddings, norms, and rope tensors are included in fixed_weight_gb.", "notes": "HF API gguf.total is 25.250987068B parameters and gguf.totalFileSize selects diffusiongemma-26B-A4B-it-BF16.gguf. A direct GGUF v3 range read found 692 tensors and 44 metadata entries. Tensor spans total 50.525009792 GB, while the linked file is 50.540834400 GB. Routed expert tensors total 45.675985920 GB and divide exactly into 128 expert indexes of 0.356843640 GB." }, "kv_adapter": { "kind": "unknown", "reason": "DiffusionGemma uses block diffusion over a 256-token canvas with a decoder that applies bidirectional attention over the generation canvas and then appends fully denoised canvases to cache. Bounds Engine v1 only models ordinary autoregressive per-output-token decode, layered KV, recurrent state, and compressed state adapters.", "notes": "The GGUF metadata confirms attention.causal false, 1024-token local sliding windows, five full-attention layers, and local/global KV head differences, but these facts are not enough to define production token throughput. A production profile needs a dedicated block-diffusion adapter with canvas length, denoising iteration count, sampler behavior, canvas self-attention traffic, cross-attention/context-cache traffic, prompt-prefix KV policy, and block append policy." }, "notes": "This profile intentionally fails closed even though GGUF metadata and tensor bytes are accessible, because the supported comparison math does not model DiffusionGemma's block-diffusion generation algorithm." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "unknown", "kv_store_bytes_per_scalar": 2, "kv_read_format": "unknown", "kv_read_bytes_per_scalar": 2, "runtime_format": "unsupported-block-diffusion-llama.cpp-gguf-bf16", "dequantization_notes": "The selected GGUF stores almost all tensor payload as BF16 plus small F32 tensors. Bounds Engine v1 does not turn those bytes into production tok/s for this repo because the generation algorithm is block diffusion rather than ordinary autoregressive decode.", "notes": "The model card says these GGUFs require the DiffusionGemma branch of llama.cpp and the dedicated llama-diffusion-cli runner; standard llama-cli and llama-server cannot generate from them yet." }, "evidence": [ { "label": "Unsloth DiffusionGemma GGUF Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/diffusiongemma-26B-A4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "selected_artifact", "downloads", "total_params_b", "max_context_tokens", "weight_format" ], "notes": "At commit f4183a2c7a354128d02545752303c4354d165bf0, the API reports a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model google/diffusiongemma-26B-A4B-it, base_model:quantized metadata, diffusion_gemma tags, region:us, 286450 downloads, gguf.total 25250987068, architecture diffusion-gemma, context_length 262144, causal false, and gguf.totalFileSize 50540834400." }, { "label": "Unsloth DiffusionGemma GGUF model card", "url": "https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "selected_artifact", "generation_algorithm", "runtime_format", "unsupported_reason" ], "notes": "The card says these GGUFs run with the DiffusionGemma branch of llama.cpp PR 24423 and require the dedicated llama-diffusion-cli runner because standard llama-cli and llama-server cannot generate from DiffusionGemma yet. It recommends Q8_0 for normal use but the HF API-selected totalFileSize points at the BF16 linked object. The card describes block-autoregressive multi-canvas sampling, iterative denoising of 256-token canvases, encoder-decoder generation, optional prompt-prefix KV cache, and 8 active / 128 total experts plus 1 shared expert." }, { "label": "Unsloth DiffusionGemma BF16 GGUF linked-object and tensor-index range read", "url": "https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF/resolve/f4183a2c7a354128d02545752303c4354d165bf0/diffusiongemma-26B-A4B-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format", "unsupported_reason" ], "notes": "HEAD checks found BF16 50.540834400 GB, Q8_0 26.878831200 GB, Q6_K 22.654490464 GB, Q5_K_M 19.145543264 GB, and Q4_K_M 16.806810208 GB. A 32MB range read of the BF16 GGUF v3 header found 692 tensors and 44 metadata entries. Tensor spans sum to 50.525009792 GB; metadata/header/alignment padding accounts for 0.015824608 GB. BF16 tensors account for 50.478940160 GB and F32 tensors for 0.046069632 GB. Non-expert tensor spans total 4.849023872 GB. Routed expert tensors total 45.675985920 GB across 30 blocks and 128 expert indexes, or 0.356843640 GB per expert index." }, { "label": "Unsloth DiffusionGemma GGUF header metadata", "url": "https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF/resolve/f4183a2c7a354128d02545752303c4354d165bf0/diffusiongemma-26B-A4B-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "generation_algorithm" ], "notes": "The GGUF metadata records general.architecture diffusion-gemma, context_length 262144, block_count 30, embedding_length 2816, head_count 16, local KV heads 8, global/full-layer KV heads 2, key/value lengths 512 for global layers, key/value lengths 256 for sliding-window layers, sliding_window 1024, five full-attention layers in the sliding-window pattern, expert_count 128, expert_used_count 8, expert_feed_forward_length 704, and attention.causal false." }, { "label": "Google DiffusionGemma 26B A4B IT base profile", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it", "source_type": "manual_review", "supports": [ "unsupported_reason", "base_model_proof" ], "notes": "The base BF16 repo is already fail-closed in Local Frontier for the same architectural reason: Bounds Engine v1 lacks a block-diffusion throughput adapter." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Reviewed from current HF API metadata, the model card, linked-object HEAD checks for all GGUF siblings, direct GGUF header/tensor-index range read of the API-selected BF16 artifact, and the existing base DiffusionGemma profile. Marked unsupported because Bounds Engine v1 lacks a block-diffusion throughput adapter." }, "unsupported_reason": "Bounds Engine v1 does not model block-diffusion generation over a denoised canvas, so ordinary autoregressive throughput would be misleading even though the selected BF16 GGUF resident weights and architecture metadata are accessible.", "notes": "This unsupported profile is a deliberate fail-closed entry. It should become audited only after a dedicated DiffusionGemma block-diffusion adapter exists." }, { "id": "unsloth--gemma-3-12b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-3-12b-it-GGUF", "title": "Unsloth Gemma 3 12B IT GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Gemma 3 12B IT.", "model_family": "gemma3-dense-gguf", "base_model_proof": { "base_model": "google/gemma-3-12b-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, package Gemma 3 config, Google base API metadata, gated base access check, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-12b-it. The Google base API is accessible for high-level metadata and confirms the Gemma license and 12187325040 BF16 parameters, but the raw base config remains gated in this audit environment. The Unsloth package config and selected GGUF header agree on the Gemma 3 12B text geometry, so this profile uses those public artifacts directly rather than claiming a direct base-config comparison." }, "architecture": { "canonical_architecture_id": "gemma-3-12b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.766034176, "swept_params_b": 11.766034176, "resident_weight_gb": 23.54015152, "swept_weight_gb": 23.533599744, "auxiliary_resident_weight_gb": 0.006551776, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-12b-it-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans; token_embd.weight is also the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for a multimodal workload", "notes": "The selected BF16 linked file is 23.540151520 GB. GGUF tensor spans total 23.533599744 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006551776 GB. The absence of output.weight means token_embd.weight remains in swept decode traffic as the tied output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 8, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 48 blocks, that yields eight full-context global layers." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 1024 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head and window metadata. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly selects a quantized KV cache." }, "notes": "Dense Gemma3 GGUF profile audited from the selected BF16 artifact and the public Unsloth Gemma 3 package config. The selected main GGUF contains text tensors only: token embedding, 48 blk.* layers, and output norm." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, image prefill, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the BF16 GGUF because gguf.totalFileSize exactly matches gemma-3-12b-it-BF16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant BF16 tensor format." }, "evidence": [ { "label": "Unsloth Gemma 3 12B IT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-3-12b-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d, the live API records a public non-gated GGUF image-text-to-text repo with base_model google/gemma-3-12b-it, base_model:quantized metadata, region:us, 97138 live downloads, GGUF architecture gemma3, context_length 131072, gguf.total 11766034176, and gguf.totalFileSize 23540151520. The catalog keeps the older qualifying scrape count 108980 until the over-100k working set is regenerated. The API totalFileSize matches gemma-3-12b-it-BF16.gguf, so this profile targets that artifact." }, { "label": "Unsloth Gemma 3 12B IT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/raw/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "multimodal_packaging" ], "notes": "The card records base_model google/gemma-3-12b-it, Gemma license, Unsloth GGUF packaging, Gemma 3 multimodal text/image behavior, 128K context for the 12B size, and no model-card override that selects a lower-bit GGUF sibling instead of the API-selected BF16 artifact." }, { "label": "Unsloth Gemma 3 12B IT package config", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/raw/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "sliding_window", "max_context_tokens", "base_model_proof" ], "notes": "The package config records Gemma3ForConditionalGeneration, gemma3_text, 48 text layers, hidden size 3840, intermediate size 15360, 16 attention heads, 8 KV heads, head_dim 256, max_position_embeddings 131072, cache_implementation hybrid, sliding_window 1024, sliding_window_pattern 6, linear RoPE scaling factor 8, vocab_size 262208, bfloat16 text dtype, and a SigLIP vision config. These fields agree with the selected GGUF header for the text artifact." }, { "label": "Google Gemma 3 12B IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-12b-it", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "The live base API records a public but manually gated Gemma-license image-text-to-text repo with base_model google/gemma-3-12b-pt, region:us, Gemma3ForConditionalGeneration high-level config metadata, and safetensors BF16 total 12187325040 parameters." }, { "label": "Google Gemma 3 12B IT gated base config access check", "url": "https://huggingface.co/google/gemma-3-12b-it/raw/96b6f1eccf38110c56df3a15bffe176da04bfd80/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The raw base config returned HTTP 401 with an access-restricted response in this audit environment. This profile therefore does not claim a direct Google base-config comparison." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 48 blocks recorded by the selected GGUF header." }, { "label": "Unsloth Gemma 3 12B IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found BF16 23.540151520 GB, Q8_0 12.510212576 GB, Q6_K 9.660811616 GB, Q5_K_M 8.445036896 GB, Q5_K_S 8.231962976 GB, Q4_K_M 7.300778336 GB, Q4_K_S 6.935333216 GB, Q3_K_M 6.008818016 GB, Q3_K_S 5.458315616 GB, Q2_K 4.768221536 GB, IQ4_NL 6.887164256 GB, UD-Q4_K_XL 7.432229216 GB, mmproj-BF16 0.854200448 GB, mmproj-F16 0.854200448 GB, and mmproj-F32 1.676341376 GB. The selected BF16 artifact exactly matches API gguf.totalFileSize; mmproj sidecars are not included in this ordinary text profile." }, { "label": "Unsloth Gemma 3 12B IT BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/resolve/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d/gemma-3-12b-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter", "max_context_tokens" ], "notes": "A 64MB range-read of the GGUF v3 header found 36 metadata entries and 626 tensors. The linked file is 23.540151520 GB. Tensor spans sum to 23.533599744 GB: token_embd.weight 2.013757440 GB, blk.* tensors 21.519826944 GB, and output_norm.weight 0.000015360 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006551776 GB. Stored tensor spans split into BF16 23.530536960 GB and F32 0.003062784 GB. The header records general.architecture gemma3, gemma3.block_count 48, context_length 131072, embedding_length 3840, feed_forward_length 15360, attention.head_count 16, attention.head_count_kv 8, attention key/value length 256, attention.sliding_window 1024, rope.freq_base 1000000, rope.scaling.factor 8, vocab_size 262208, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, Google base API metadata, package config, gated-base access check, selected linked-object HEAD checks, existing Gemma 3 layered-KV treatment, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the API-selected Gemma 3 12B IT BF16 GGUF artifact. Do not infer lower-bit sibling footprints or multimodal projector residency unless the workload explicitly selects and audits those files." }, { "id": "unsloth--gemma-3-270m-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-3-270m-it-GGUF", "title": "Unsloth Gemma 3 270M IT GGUF F16", "summary": "Audited memory-side text-decode bounds profile for the API-selected F16 GGUF artifact of Gemma 3 270M IT.", "model_family": "gemma3-dense-gguf", "base_model_proof": { "base_model": "google/gemma-3-270m-it", "relation": "quantized", "source": "Hugging Face model card/API metadata, Google base API metadata, gated base config access check, package side-file review, and selected GGUF header metadata", "config_compatible": false, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-3-270m-it. The Google base API is accessible for high-level metadata and confirms Gemma3ForCausalLM, Gemma license, 32768-token output limit in the model card, and 268098176 BF16 parameters, but raw base config access remains gated in this audit environment. This profile therefore uses the selected public GGUF header as the architecture source." }, "architecture": { "canonical_architecture_id": "gemma-3-270m", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.268098176, "swept_params_b": 0.268098176, "resident_weight_gb": 0.542835488, "swept_weight_gb": 0.536308224, "auxiliary_resident_weight_gb": 0.006527264, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-3-270m-it-F16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans; token_embd.weight is also the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The selected F16 linked file is 0.542835488 GB. GGUF tensor spans total 0.536308224 GB, while GGUF metadata, tokenizer, header, and alignment padding account for 0.006527264 GB. The absence of output.weight means token_embd.weight remains in swept decode traffic as the tied output projection." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 3, "kv_heads": 1, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Gemma 3 uses five local attention layers between global attention layers. For 18 blocks, that yields three full-context global layers." }, { "kind": "sliding_window", "layers": 15, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata records gemma3.attention.sliding_window 512 for local layers." } ], "notes": "Layered KV uses the Gemma 3 local/global pattern and the selected GGUF artifact's own KV head and window metadata." }, "notes": "Dense Gemma3 GGUF profile audited from the selected F16 artifact, not from the gated Google raw config." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-f16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected F16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, activation traffic, and compute throughput are outside Bounds Engine v1.", "notes": "The API-selected artifact is the F16 GGUF because gguf.totalFileSize exactly matches gemma-3-270m-it-F16.gguf. Explicit resident and swept byte fields are authoritative; the weight_bytes_per_param field identifies the dominant F16 tensor format." }, "evidence": [ { "label": "Unsloth Gemma 3 270M IT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-3-270m-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit c90975dbd40c0c7b275fefaae758c3415c906238, the live API records a public non-gated GGUF text-generation repo with base_model google/gemma-3-270m-it, base_model:quantized metadata, Gemma license, region:us, 105716 downloads, GGUF architecture gemma3, context_length 32768, gguf.total 268098176, and gguf.totalFileSize 542835488. The API totalFileSize matches gemma-3-270m-it-F16.gguf, so this profile targets that artifact." }, { "label": "Unsloth Gemma 3 270M IT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "runtime_format" ], "notes": "The card metadata records base_model google/gemma-3-270m-it, Gemma license, text-generation pipeline, transformers library metadata, and GGUF packaging. The visible card includes Gemma 3 270M guidance and repeats the 32K-token context/output limit for the 270M and 1B sizes." }, { "label": "Google Gemma 3 270M IT API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-270m-it", "source_type": "model_card", "supports": [ "base_model_proof", "license", "total_params_b" ], "notes": "The live base API records a manually gated Gemma-license text-generation repo with Gemma3ForCausalLM high-level config metadata, base_model google/gemma-3-270m, region:us, and safetensors BF16 total 268098176 parameters. Direct raw config access remains blocked." }, { "label": "Google Gemma 3 270M IT gated config access check", "url": "https://huggingface.co/google/gemma-3-270m-it/raw/ac82b4e820549b854eebf28ce6dedaf9fdfa17b3/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Raw config access returned HTTP 401 with an access-restricted response, and hf download with the configured CLI token returned access denied because the repository requires approval. The existing google/gemma-3-270m-it profile therefore remains an unsupported gated-base stub, and this GGUF profile does not claim a direct base-config comparison." }, { "label": "Unsloth Gemma 3 270M IT GGUF side files", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF/tree/c90975dbd40c0c7b275fefaae758c3415c906238", "source_type": "manual_review", "supports": [ "serving", "max_context_tokens" ], "notes": "The package side files params and template were downloaded at the pinned commit. params sets num_predict 32768 and stop tokens for llama.cpp/Ollama-style serving; template is the Gemma start_of_turn chat template. The side files do not override the selected GGUF architecture metadata." }, { "label": "Gemma 3 local/global attention description", "url": "https://developers.googleblog.com/gemma-explained-whats-new-in-gemma-3/", "source_type": "vendor_doc", "supports": [ "kv_adapter", "layer_pattern" ], "notes": "Google's Gemma 3 architecture note describes repeating interleaving blocks with five local attention layers and one global attention layer. This profile applies that pattern to the 18 blocks recorded by the selected GGUF header." }, { "label": "Unsloth Gemma 3 270M IT GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF/tree/c90975dbd40c0c7b275fefaae758c3415c906238", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found F16 0.542835488 GB, Q8_0 0.291546144 GB, Q4_K_M 0.253115424 GB, Q2_K 0.237079584 GB, UD-Q4_K_XL 0.253934624 GB, UD-IQ2_XXS 0.180104224 GB, and imatrix_unsloth.gguf_file 0.000471040 GB. The selected F16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Gemma 3 270M IT F16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF/resolve/c90975dbd40c0c7b275fefaae758c3415c906238/gemma-3-270m-it-F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "max_context_tokens" ], "notes": "A 16MB range-read of the GGUF v3 header found 41 metadata entries and 236 tensors. The linked file is 0.542835488 GB. Tensor spans sum to 0.536308224 GB: token_embd.weight 0.335544320 GB, blk.* tensors 0.200761344 GB, and output_norm.weight 0.000002560 GB. Metadata/tokenizer/header/alignment overhead accounts for 0.006527264 GB. Stored tensor spans split into F16 0.536084480 GB and F32 0.000223744 GB. The header records general.architecture gemma3, gemma3.block_count 18, context_length 32768, embedding_length 640, feed_forward_length 2048, attention.head_count 4, attention.head_count_kv 1, key/value length 256, attention.sliding_window 512, rope.freq_base 1000000, vocab size 262144, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card metadata, Google base API metadata, gated-base config access check, package side-file review, selected linked-object HEAD checks, Google Gemma 3 local/global attention documentation, and a direct GGUF header/tensor-index range read of the selected F16 artifact." }, "notes": "Use this profile for the API-selected Gemma 3 270M IT F16 GGUF artifact. Do not infer the lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "unsloth--gemma-4-12b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-12b-it-GGUF", "title": "Unsloth Gemma 4 12B IT GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Gemma 4 12B IT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF CLI/API GGUF metadata, Google base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it. The selected GGUF header records the same Gemma 4 12B text geometry as the Google config. The Unsloth repo does not ship config.json, so the immutable Google config is used for the high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 23.832065184, "swept_weight_gb": 23.816242688, "auxiliary_resident_weight_gb": 0.015822496, "resident_parameter_scope": "selected GGUF linked file size for gemma-4-12b-it-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; separate mmproj and MTP GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 23.816242688 GB, while the linked file size is 23.832065184 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact. Audio, image, video projection, and MTP/speculative sidecars are separate files and require separate workload profiles if loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param records the BF16 tensor storage format." }, "evidence": [ { "label": "Unsloth Gemma 4 12B GGUF HF CLI/API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-12b-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF CLI/API response at commit 3249fa54d5efa384afc552cc6700ad091efd5c39 records base_model google/gemma-4-12B-it, downloads 1443180, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 23832065184." }, { "label": "Unsloth Gemma 4 12B GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "attention_pattern", "max_context_tokens", "unified_multimodal" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-12B-it, and the Google Gemma 4 architecture description: 12B Unified, encoder-free multimodal projection, and up to 256K context." }, { "label": "Google Gemma 4 12B IT config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "Unsloth Gemma 4 12B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/tree/3249fa54d5efa384afc552cc6700ad091efd5c39", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found gemma-4-12b-it-BF16.gguf has linked size 23832065184 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files, mmproj-BF16/F16/F32.gguf, mtp-gemma-4-12b-it.gguf, and MTP/* sidecar GGUFs have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Gemma 4 12B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/resolve/3249fa54d5efa384afc552cc6700ad091efd5c39/gemma-4-12b-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 54 metadata entries and 667 tensors. The linked file is 23.832065184 GB. Tensor spans sum to 23.816242688 GB: token_embd.weight 2.01326592 GB, blk.* tensors 21.802960384 GB, output_norm.weight 0.00001536 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/alignment bytes account for 0.015822496 GB. Actual tensor bytes are 23.816241344 GB, with 1344 bytes of tensor-alignment padding. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the live HF CLI/API metadata, model card, immutable Google config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--gemma-4-12b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-12B-it-qat-GGUF", "title": "Unsloth Gemma 4 12B IT QAT GGUF UD-Q4_K_XL", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth UD-Q4_K_XL GGUF artifact of Gemma 4 12B IT QAT.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google QAT unquantized config, linked-object size checks, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-12B-it-qat-q4_0-unquantized. The selected GGUF header records the same Gemma 4 12B text geometry as the Google QAT unquantized config. The Unsloth GGUF repo does not ship config.json, so the immutable Google QAT unquantized config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 6.716355328, "swept_weight_gb": 6.700533248, "auxiliary_resident_weight_gb": 0.01582208, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-12B-it-qat-UD-Q4_K_XL.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as model tensors; mmproj and MTP GGUF files are separate sidecars and are not included unless explicitly loaded for another workload", "notes": "The profile targets the UD-Q4_K_XL GGUF file selected by the model card's llama.cpp command and by HF API gguf.totalFileSize. Header tensor spans total 6.700533248 GB, while the linked file size is 6.716355328 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary target-model text decode for the selected main UD-Q4_K_XL GGUF artifact after any multimodal prefill. Multimodal projector and MTP/speculative sidecar files are separate artifacts and should get separate workload profiles if they are loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5640511955310384, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-q4-k-xl-qat-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, speculative MTP drafter execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth UD-Q4_K_XL GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Unsloth Gemma 4 12B IT QAT GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-12B-it-qat-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 7102bdea62863acff919c945405ef29973113d66 records base_model google/gemma-4-12B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, any-to-any pipeline, region:us, 509756 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 6716355328." }, { "label": "Unsloth Gemma 4 12B IT QAT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF/raw/7102bdea62863acff919c945405ef29973113d66/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "speculative_sidecar", "max_context_tokens" ], "notes": "The card records base_model google/gemma-4-12B-it-qat-q4_0-unquantized and points llama.cpp serving to unsloth/gemma-4-12B-it-qat-GGUF:UD-Q4_K_XL. It also states that the repo ships a separate MTP drafter at mtp-gemma-4-12B-it.gguf, which is deliberately excluded from this ordinary target-model profile." }, { "label": "Google Gemma 4 12B IT QAT unquantized config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "Unsloth Gemma 4 12B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF/tree/7102bdea62863acff919c945405ef29973113d66", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope", "sidecar_files" ], "notes": "HEAD checks found gemma-4-12B-it-qat-UD-Q4_K_XL.gguf is 6716355328 bytes, exactly matching API gguf.totalFileSize. Separate sidecars are mmproj-BF16/F16 175115840 bytes each, mmproj-F32 209522240 bytes, mtp-gemma-4-12B-it.gguf 253707328 bytes, MTP Q4_0 253707328 bytes, MTP Q8_0 465126464 bytes, and MTP BF16/F16 861537344 bytes each." }, { "label": "Unsloth Gemma 4 12B IT QAT UD-Q4_K_XL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF/resolve/7102bdea62863acff919c945405ef29973113d66/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 48 metadata entries and 667 tensors. The selected file is 6.716355328 GB, with tensor payloads starting at byte 15822080. Tensor spans total 6.700533248 GB across 11907350576 logical elements: token_embd.weight 0.566231040 GB, blk.* tensors 6.134285824 GB, output_norm.weight 0.000015360 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_0 6.697451520 GB and F32 0.003081728 GB. Metadata/tokenizer/header/file overhead accounts for 0.015822080 GB. The header records general.name Gemma-4 12B IT (smart Q4_0, QAT-lossless), general.file_type 2, general.quantized_by unsloth, gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, pinned model card, immutable Google QAT unquantized config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth main UD-Q4_K_XL target-model text artifact. Do not infer multimodal projector or MTP drafter residency from the repo name or the -hf convenience command; those sidecars need explicit workload profiles if loaded." }, { "id": "unsloth--gemma-4-26b-a4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "title": "Unsloth Gemma 4 26B A4B GGUF MXFP4 MoE", "summary": "Audited memory-side bounds profile for the selected Unsloth MXFP4_MOE GGUF artifact of Gemma 4 26B A4B.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it", "relation": "quantized", "source": "Hugging Face model card base_model metadata, Unsloth GGUF API metadata, config.json, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of google/gemma-4-26B-A4B-it. The GGUF and config metadata record the same Gemma 4 26B A4B text architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 16.551046944, "main_resident_weight_gb": 16.535224256, "auxiliary_resident_weight_gb": 0.015822688, "fixed_weight_gb": 2.578677696, "routed_expert_weight_gb": 0.10903552, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for gemma-4-26B-A4B-it-MXFP4_MOE.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main GGUF artifact", "auxiliary_scope": "GGUF metadata, header, and alignment padding are resident in the selected artifact but not swept as model tensors; separate mmproj and MTP GGUF files in the repo are not included unless explicitly loaded for another workload", "shared_expert_notes": "The card states 8 active / 128 total experts and 1 shared expert. The GGUF header stores dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight; those always-on/shared tensors are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects gemma-4-26B-A4B-it-MXFP4_MOE.gguf. A GGUF v3 range-read found 658 tensors. Tensor spans total 16.535224256 GB, while the linked file is 16.551046944 GB. Routed expert tensors total 13.95654656 GB and divide exactly by 128 expert indexes. The final layer uses larger expert tensor classes than layers 0-28, so this profile uses the exact total routed expert bytes rather than a per-layer rounded estimate." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The config layer_types and GGUF head-count array show five full-attention layers. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile targets the selected main MXFP4_MOE text GGUF artifact. Multimodal projector and speculative/MTP sidecar files are separate artifacts and should get separate workload profiles if they are loaded." }, "serving": { "weight_format": "mxfp4", "weight_bytes_per_param": 0.655924931, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth MXFP4_MOE GGUF. Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Gemma 4 26B A4B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-26B-A4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The API response at commit 3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a records base_model google/gemma-4-26B-A4B-it, GGUF architecture gemma4, 262144 context length, gguf.total 25233142046, and gguf.totalFileSize 16551046944." }, { "label": "Unsloth Gemma 4 26B A4B GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The card records base_model google/gemma-4-26B-A4B-it and includes the Google Gemma 4 architecture table: 25.2B total, 3.8B active, 30 layers, 1024-token sliding window, 256K context, and 8 active / 128 total experts plus 1 shared expert." }, { "label": "Unsloth Gemma 4 26B A4B GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/raw/3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, bfloat16 source dtype, 30 text layers, 16 attention heads, 8 local KV heads, 5 full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, and 262144 max position embeddings." }, { "label": "Unsloth Gemma 4 MXFP4_MOE GGUF linked-object and tensor-index range read", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/resolve/3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a/gemma-4-26B-A4B-it-MXFP4_MOE.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "HF linked-object metadata reports 16.551046944 GB for gemma-4-26B-A4B-it-MXFP4_MOE.gguf, matching API gguf.totalFileSize. A 64MB range-read of the GGUF v3 header found 658 tensors and 60 metadata entries. Tensor spans sum to 16.535224256 GB; metadata/header/alignment padding accounts for 0.015822688 GB. Non-expert tensor spans total 2.578677696 GB. Routed expert tensors total 13.95654656 GB across 30 layers and 128 expert indexes, or 0.10903552 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, unified full-attention K/V geometry, and separate sliding-layer K/V projections." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the live HF API, model card, config.json, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected MXFP4_MOE artifact." }, "notes": "Use this profile for the Unsloth main GGUF text artifact. Do not infer it from the base BF16 Google repo, the NVIDIA ModelOpt NVFP4 repo, or the separate mmproj/MTP files." }, { "id": "unsloth--gemma-4-26b-a4b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-26B-A4B-it-qat-GGUF", "title": "Unsloth Gemma 4 26B A4B QAT GGUF Q4_0 MoE", "summary": "Audited memory-side bounds profile for the selected Unsloth UD-Q4_K_XL GGUF artifact of the QAT Gemma 4 26B A4B model.", "model_family": "gemma4-moe", "base_model_proof": { "base_model": "google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card base_model metadata, Unsloth GGUF API metadata, repo config.json, base repo config.json, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of google/gemma-4-26B-A4B-it-qat-q4_0-unquantized. The packaged config, base config, model card, and GGUF header record matching Gemma 4 26B A4B text architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-26b-a4b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 14.24904512, "main_resident_weight_gb": 14.233223104, "auxiliary_resident_weight_gb": 0.015822016, "fixed_weight_gb": 1.386856384, "routed_expert_weight_gb": 0.10036224, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GGUF linked file size for gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf", "traffic_scope": "ordinary text decode charges all non-expert GGUF tensor spans plus expected distinct routed expert tensor spans from the selected main text GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors; separate mmproj and MTP GGUF files in the repo are not included unless explicitly loaded for another workload", "shared_expert_notes": "The model card records 8 active / 128 total experts and 1 shared expert. The GGUF header stores dense blk.*.ffn_down/gate/up tensors outside blk.*.ffn_*_exps.weight; those always-on/shared tensors are charged in fixed_weight_gb.", "notes": "The HF API gguf.total is 25.233142046B parameters and gguf.totalFileSize selects gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf. A GGUF v3 range-read found 658 tensors and 50 metadata entries. Tensor spans total 14.233223104 GB, while the linked file is 14.249045120 GB. Routed expert tensors are the 30 ffn_down_exps and 30 ffn_gate_up_exps Q4_0 tensors; they total 12.846366720 GB and divide exactly by 128 expert indexes. Non-expert tensor spans total 1.386856384 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 5, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "The config layer_types and GGUF head-count array show five full-attention layers. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 25, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 25 layers use 1024-token local sliding-window attention with 8 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile targets the selected main QAT UD-Q4_K_XL text GGUF artifact. Multimodal projector and speculative/MTP sidecar files are separate artifacts and should get separate workload profiles if they are loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.564695633, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4_0-qat-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth QAT UD-Q4_K_XL GGUF. The GGUF header records general.file_type 2, general.quantized_by unsloth, and general.name Gemma-4 26B-A4B IT (smart Q4_0, QAT-lossless). Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Gemma 4 26B A4B QAT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-26B-A4B-it-qat-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The API response at commit 02749a7b272109255a4c559a80894d3d9777574c records base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, GGUF architecture gemma4, 262144 context length, gguf.total 25233142046, and gguf.totalFileSize 14249045120." }, { "label": "Unsloth Gemma 4 26B A4B QAT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF/raw/02749a7b272109255a4c559a80894d3d9777574c/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "attention_pattern", "max_context_tokens" ], "notes": "The card records base_model google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, points llama.cpp to unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL, and includes the Google Gemma 4 architecture table: 25.2B total, 3.8B active, 30 layers, 1024-token sliding window, 256K context, and 8 active / 128 total experts plus 1 shared expert." }, { "label": "Unsloth Gemma 4 26B A4B QAT GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF/raw/02749a7b272109255a4c559a80894d3d9777574c/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "top_k_experts", "routed_experts", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, bfloat16 source dtype, 30 text layers, 16 attention heads, 8 local KV heads, 5 full-attention layers, 25 sliding-attention layers, 1024-token sliding window, num_experts 128, top_k_experts 8, attention_k_eq_v true, tie_word_embeddings true, and 262144 max position embeddings." }, { "label": "Google Gemma 4 26B A4B QAT unquantized base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "routed_experts", "routed_experts_per_token" ], "notes": "The base config records the same Gemma 4 text geometry: 30 layers, 16 attention heads, 8 local KV heads, 2 global KV heads, 1024-token sliding window, five full-attention layers, 128 experts, top_k_experts 8, full-attention K=V behavior, tied embeddings, and 262144 max positions." }, { "label": "Unsloth QAT UD-Q4_K_XL GGUF linked-object and tensor-index range read", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF/resolve/02749a7b272109255a4c559a80894d3d9777574c/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "HF linked-object metadata reports 14.249045120 GB for gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf, matching API gguf.totalFileSize. A 64MB range-read of the GGUF v3 header found 658 tensors and 50 metadata entries. Tensor spans sum to 14.233223104 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015822016 GB. Non-expert tensor spans total 1.386856384 GB. Routed expert tensors total 12.846366720 GB across 30 layers and 128 expert indexes, or 0.100362240 GB per expert index. The header records gemma4.block_count 30, context_length 262144, expert_count 128, expert_used_count 8, sliding_window 1024, local KV heads 8, full-layer KV heads 2, unified full-attention K/V geometry, and separate sliding-layer K/V projections." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from the live HF API, pinned model card, repo config.json, base config.json, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth main QAT GGUF text artifact. Do not infer it from the generated metadata estimate, the non-QAT MXFP4_MOE GGUF profile, or the separate mmproj/MTP files." }, { "id": "unsloth--gemma-4-31b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-31B-it-GGUF", "title": "Unsloth Gemma 4 31B IT GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth IQ4_NL GGUF artifact of Gemma 4 31B IT.", "model_family": "gemma4-dense-multimodal", "base_model_proof": { "base_model": "google/gemma-4-31B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Unsloth config, Google base config, linked-object sizes, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-31B-it. The Unsloth config and selected GGUF header record the same Gemma 4 31B dense text geometry as the Google config." }, "architecture": { "canonical_architecture_id": "gemma-4-31b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 17.287668736, "swept_weight_gb": 17.271836544, "auxiliary_resident_weight_gb": 0.015832192, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-31B-it-IQ4_NL.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; BF16 shards, mmproj sidecars, and MTP GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the IQ4_NL GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 17.271836544 GB, while the linked file size is 17.287668736 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no output.weight, mmproj, vision, audio, MTP, or draft tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, 47, 53, and 59. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_NL GGUF artifact after any multimodal prefill. The BF16 shards, multimodal projector sidecars, and MTP/speculative sidecars are separate files and require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.56316493821709, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth IQ4_NL GGUF because HF API gguf.totalFileSize matches gemma-4-31B-it-IQ4_NL.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Unsloth Gemma 4 31B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-31B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 8906b3db2e669a0b1d6293c315d3f9fbf934a86d records base_model google/gemma-4-31B-it, Apache-2.0 license, public image-text-to-text GGUF repo, region:us, imatrix metadata, 607972 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, and gguf.totalFileSize 17287668736." }, { "label": "Unsloth Gemma 4 31B GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "attention_pattern", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-31B-it, MTP sidecar support added June 9, and the Google Gemma 4 31B dense architecture table: 30.7B parameters, 60 layers, 1024-token sliding window, 256K context, and text/image modalities." }, { "label": "Unsloth Gemma 4 31B GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/raw/8906b3db2e669a0b1d6293c315d3f9fbf934a86d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The pinned Unsloth config records Gemma4ForConditionalGeneration, bfloat16 source dtype, tie_word_embeddings true, 60 text layers, ten full-attention layers, 50 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, four global KV heads, 16 local KV heads, 512 global key/value length, 256 local key/value length, resident vision config, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B IT base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching audited text_config and vision_config geometry fields between the Google BF16 repo and this Unsloth GGUF package." }, { "label": "Unsloth Gemma 4 31B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/tree/8906b3db2e669a0b1d6293c315d3f9fbf934a86d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "The repo contains BF16 shards totaling 61.413188544 GB, quantized GGUF siblings from 8.534293504 GB to 35.020039168 GB, MTP sidecars from 0.514687104 GB to 0.954843264 GB, and mmproj sidecars from 1.198957024 GB to 2.30299696 GB. gemma-4-31B-it-IQ4_NL.gguf has linked size 17.287668736 GB, exactly matching API gguf.totalFileSize, and is the selected main artifact for this profile." }, { "label": "Unsloth Gemma 4 31B IQ4_NL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/resolve/8906b3db2e669a0b1d6293c315d3f9fbf934a86d/gemma-4-31B-it-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 57 metadata entries and 833 tensors. The selected file is 17.287668736 GB, with tensor payloads starting at byte 15832192. Tensor spans total 17.271836544 GB across 30697345596 logical elements: token_embd.weight 0.792723456 GB, blk.* tensors 16.47909056 GB, output_norm.weight 0.000021504 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into IQ4_NL 16.47378432 GB, Q4_K 0.792723456 GB, and F32 0.005328768 GB. Metadata/tokenizer/header/file overhead accounts for 0.015832192 GB. The header records gemma4.block_count 60, context_length 262144, attention.head_count 32, layer KV head array with ten full layers using four KV heads and 50 sliding layers using 16 KV heads, key/value length 512 for full layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, pinned Unsloth config, Google base config comparison, linked-object file sizes, and a direct GGUF header/tensor-index range read of the selected IQ4_NL artifact." }, "notes": "Use this profile for the Unsloth main IQ4_NL GGUF text artifact. Do not infer BF16 shard residency, multimodal projector residency, or MTP sidecar traffic unless those separate files are explicitly loaded by the workload." }, { "id": "unsloth--gemma-4-31b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-31B-it-qat-GGUF", "title": "Unsloth Gemma 4 31B QAT GGUF Q4_0 Dense", "summary": "Audited memory-side bounds profile for the selected Unsloth UD-Q4_K_XL GGUF artifact of the QAT Gemma 4 31B dense model.", "model_family": "gemma4-dense", "base_model_proof": { "base_model": "google/gemma-4-31B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card base_model metadata, Unsloth GGUF API metadata, repo config.json, base repo config.json, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of google/gemma-4-31B-it-qat-q4_0-unquantized. The packaged config, base config, model card, and GGUF header record matching Gemma 4 31B dense text architecture." }, "architecture": { "canonical_architecture_id": "gemma-4-31b-qat", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 30.697345596, "swept_params_b": 30.697345596, "auxiliary_resident_params_b": 0, "resident_weight_gb": 17.287668064, "swept_weight_gb": 17.271834864, "auxiliary_resident_weight_gb": 0.0158332, "resident_parameter_scope": "selected GGUF linked file size and header tensor elements for gemma-4-31B-it-qat-UD-Q4_K_XL.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main artifact; token_embd.weight is charged as tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer data, header, and alignment padding are resident in the selected artifact but not swept as model tensors; separate mmproj and MTP GGUF files in the repo are not included unless explicitly loaded for another workload", "notes": "The HF API gguf.total is 30.697345596B parameters and gguf.totalFileSize selects gemma-4-31B-it-qat-UD-Q4_K_XL.gguf. A GGUF v3 range-read found 833 tensors and 47 metadata entries. Tensor spans total 17.271834864 GB, while the linked file is 17.287668064 GB. The main GGUF has no output.weight tensor, so token_embd.weight is tied and remains swept as output-projection traffic for ordinary decode. The selected main file has no mmproj, vision, visual, audio, or MTP tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 4, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "The config layer_types and GGUF head-count array show ten full-attention layers. Official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 50, "kv_heads": 16, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 50 layers use 1024-token local sliding-window attention with 16 KV heads. Official Gemma4Attention reuses K as V only for non-sliding full-attention layers; sliding layers instantiate separate k_proj and v_proj tensors, so this component charges separate K and V streams." } ], "notes": "Hybrid local/global attention is represented as an explicit layered KV sum. The profile uses FP16 KV because the GGUF package does not declare a required quantized KV cache format. Full-attention layers use the Gemma 4 K=V path; sliding-window layers keep separate K and V caches." }, "notes": "This profile targets the selected main QAT UD-Q4_K_XL text GGUF artifact. Multimodal projector and speculative/MTP sidecar files are separate artifacts and should get separate workload profiles if they are loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.563164916, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4_0-qat-dense-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth QAT UD-Q4_K_XL GGUF. The GGUF header records general.file_type 2, general.quantized_by unsloth, and general.name Gemma-4 31B IT (smart Q4_0, QAT-lossless). Default GGUF KV is modeled as FP16 unless the serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Gemma 4 31B QAT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-31B-it-qat-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The API response at commit 365d657136993b4d7c40d868dd45ecb7a48e7ebf records base_model google/gemma-4-31B-it-qat-q4_0-unquantized, GGUF architecture gemma4, 262144 context length, gguf.total 30697345596, gguf.totalFileSize 17287668064, and 407993 current downloads." }, { "label": "Unsloth Gemma 4 31B QAT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/raw/365d657136993b4d7c40d868dd45ecb7a48e7ebf/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "selected_artifact", "attention_pattern", "max_context_tokens", "sidecar_artifacts" ], "notes": "The card records base_model google/gemma-4-31B-it-qat-q4_0-unquantized, points llama.cpp to unsloth/gemma-4-31B-it-qat-GGUF:UD-Q4_K_XL, documents that a root MTP drafter can be auto-discovered separately, and includes the Google Gemma 4 architecture table: 30.7B dense parameters, 60 layers, 1024-token sliding window, 256K context, 262K vocabulary, text/image modalities, and about 550M vision encoder parameters." }, { "label": "Unsloth Gemma 4 31B QAT GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/raw/365d657136993b4d7c40d868dd45ecb7a48e7ebf/config.json", "source_type": "config", "supports": [ "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v", "tie_word_embeddings", "max_context_tokens" ], "notes": "The config records Gemma4ForConditionalGeneration, BF16 source dtype, 60 text layers, 32 attention heads, 16 local KV heads, 4 global KV heads, 50 sliding-attention layers, 10 full-attention layers, 1024-token sliding window, dense text MLPs with enable_moe_block false, attention_k_eq_v true, tie_word_embeddings true, and 262144 max position embeddings." }, { "label": "Google Gemma 4 31B QAT unquantized base config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "attention_k_eq_v" ], "notes": "The base config records the same Gemma 4 31B text geometry used by this profile: 60 layers, 32 attention heads, 16 local KV heads, 4 global KV heads, 1024-token sliding window, ten full-attention layers, dense MLPs, full-attention K=V behavior, tied embeddings, and 262144 max positions." }, { "label": "Unsloth QAT UD-Q4_K_XL GGUF linked-object and tensor-index range read", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/resolve/365d657136993b4d7c40d868dd45ecb7a48e7ebf/gemma-4-31B-it-qat-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "weight_format", "kv_adapter", "sidecar_artifacts" ], "notes": "HF linked-object metadata reports 17.287668064 GB for gemma-4-31B-it-qat-UD-Q4_K_XL.gguf, matching API gguf.totalFileSize. A 128MB range-read of the GGUF v3 header found 833 tensors and 47 metadata entries. Tensor spans sum to 17.271834864 GB; metadata/header/tokenizer/alignment overhead accounts for 0.015833200 GB. Tensor spans split into Q4_0 17.266507776 GB and F32 0.005327088 GB. token_embd.weight is Q4_0 with shape [5376, 262144] and contributes 0.792723456 GB. The selected main file has no output.weight, mmproj, vision, visual, audio, or MTP tensor. The header records gemma4.block_count 60, context_length 262144, sliding_window 1024, local KV head counts 16, full-layer KV head counts 4, key/value length 512, shared_kv_layers 0, and K=V full-attention geometry." }, { "label": "Unsloth Gemma 4 31B QAT GGUF sidecar linked-object checks", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/tree/365d657136993b4d7c40d868dd45ecb7a48e7ebf", "source_type": "derived_calculation", "supports": [ "sidecar_artifacts", "resident_weight_gb" ], "notes": "HEAD checks found separate sidecars: mmproj-BF16 1.200726496 GB, mmproj-F16 1.198957024 GB, mmproj-F32 2.302996960 GB, root mtp-gemma-4-31B-it 0.279954368 GB, MTP BF16 0.954860480 GB, MTP F16 0.954860480 GB, MTP Q4_0 0.279954368 GB, and MTP Q8_0 0.514704320 GB. These are not included in the ordinary main-artifact text profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from the live HF API, pinned model card, repo config.json, base config.json, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth main QAT GGUF text artifact. Do not include multimodal projector or speculative MTP sidecar residency unless the workload explicitly selects those files." }, { "id": "unsloth--gemma-4-e2b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-E2B-it-GGUF", "title": "Unsloth Gemma 4 E2B IT GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Gemma 4 E2B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google base config, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E2B-it. The selected GGUF header records the same Gemma 4 E2B text geometry as the Google config. The Unsloth repo does not ship config.json, so the immutable Google config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.647450147, "swept_params_b": 2.298639907, "auxiliary_resident_params_b": 2.34881024, "resident_weight_gb": 9.311303552, "swept_weight_gb": 4.597865568, "auxiliary_resident_weight_gb": 4.713437984, "resident_parameter_scope": "selected GGUF linked file size for gemma-4-E2B-it-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; separate mmproj and MTP GGUF files are not included unless explicitly loaded for another workload", "notes": "Gemma 4 E2B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. The Google config records global_head_dim 512 and attention_k_eq_v false, and the selected GGUF header contains separate attn_k and attn_v tensors for every layer." }, { "kind": "sliding_window", "layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj and MTP sidecars require separate workload profiles if loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF because HF API gguf.totalFileSize matches gemma-4-E2B-it-BF16.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param records the BF16 tensor storage format." }, "evidence": [ { "label": "Unsloth Gemma 4 E2B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E2B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8 records base_model google/gemma-4-E2B-it, Apache-2.0 license, image-text-to-text pipeline, region:us, 608014 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 4647450147, and gguf.totalFileSize 9311303552." }, { "label": "Unsloth Gemma 4 E2B GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-E2B-it, and an Unsloth GGUF package with main text, mmproj, and MTP sidecar files." }, { "label": "Google Gemma 4 E2B IT config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, one KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "Unsloth Gemma 4 E2B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF/tree/ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E2B-it-BF16.gguf is 9311303552 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files, mmproj-BF16/F16/F32.gguf, mtp-gemma-4-E2B-it.gguf, and MTP/* sidecar GGUFs have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Gemma 4 E2B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF/resolve/ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8/gemma-4-E2B-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 52 metadata entries and 601 tensors. The linked file is 9.311303552 GB. Tensor spans sum to 9.295486048 GB: per_layer_token_embd.weight 4.69762048 GB / 2348810240 logical elements, token_embd.weight 0.805306368 GB, block tensors 3.765025888 GB, per_layer_model_proj/proj_norm tensors 0.027526144 GB, output_norm.weight 0.000006144 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/alignment bytes account for 0.015817504 GB. Swept tensor spans excluding the per-layer token lookup table total 4.597865568 GB / 2298639907 logical elements. The header records gemma4.block_count 35, context_length 131072, attention.head_count 8, attention.head_count_kv 1, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, embedding_length_per_layer_input 256, and separate attn_k and attn_v tensors for every layer. The selected main file has no mmproj, vision, audio, MTP, or output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google config, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--gemma-4-e2b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-E2B-it-qat-GGUF", "title": "Unsloth Gemma 4 E2B IT QAT GGUF UD-Q2_K_XL", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth UD-Q2_K_XL GGUF artifact of Gemma 4 E2B IT QAT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E2B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, repo config.json, Google QAT unquantized config metadata, existing Gemma 4 E2B QAT profile comparison, Transformers PLE documentation, linked-object HEAD checks, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of google/gemma-4-E2B-it-qat-q4_0-unquantized. The packaged config and selected GGUF header record the same Gemma 4 E2B text geometry as the already audited Google QAT profile: 35 text layers, seven full-attention layers, 28 sliding-attention layers, shared K/V across the final 20 layers, tied embeddings, PLE tensors, and 131072 max position embeddings." }, "architecture": { "canonical_architecture_id": "gemma-4-e2b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.628569635, "swept_params_b": 2.279759395, "auxiliary_resident_params_b": 2.34881024, "resident_weight_gb": 2.186184768, "swept_weight_gb": 0.849165408, "auxiliary_resident_weight_gb": 1.33701936, "resident_parameter_scope": "selected GGUF linked file size and API GGUF logical tensor parameters for gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as full matrices per generated token; mmproj and MTP sidecar GGUF files are not included unless explicitly loaded for another workload", "notes": "Gemma 4 E2B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection linear layer. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per_layer_proj_norm.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "alloc_layers": 3, "read_layers": 7, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 4, 9, 14, 19, 24, 29, and 34. The config and GGUF header record shared_kv_layers 20, so only layers 4, 9, and 14 allocate full-context K/V cache, while all seven full-attention layers read full-context K/V. The config records global_head_dim 512 and attention_k_eq_v false." }, { "kind": "sliding_window", "layers": 28, "alloc_layers": 12, "read_layers": 28, "kv_heads": 1, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records shared_kv_layers 20. Only the first twelve sliding-attention layers allocate K/V cache, while all 28 sliding-attention layers read up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. Gemma 4 E2B shares K/V across the final 20 decoder layers, so allocation layer counts differ from read layer counts. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main UD-Q2_K_XL GGUF artifact after any multimodal prefill. Separate mmproj and MTP sidecars require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.47232405265520006, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-q2-k-xl-qat-ple-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, MTP sidecar execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth UD-Q2_K_XL GGUF because HF API gguf.totalFileSize matches gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf. The GGUF header stores TQ2_0, Q4_0, Q8_0, F16, and F32 tensor spans. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Unsloth Gemma 4 E2B IT QAT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E2B-it-qat-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit db01ae3ceeca98487bf3569814f832f5023cd48c records base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, Apache-2.0 license metadata, any-to-any pipeline, region:us, 135267 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 4628569635, and gguf.totalFileSize 2186184768." }, { "label": "Unsloth Gemma 4 E2B IT QAT GGUF model card metadata", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "model_family" ], "notes": "The card metadata records Apache-2.0 licensing, base_model google/gemma-4-E2B-it-qat-q4_0-unquantized, and Unsloth Gemma/Google tags. The README body is empty at the audited revision, so source claims come from API metadata, config metadata, and direct file/header inspection." }, { "label": "Unsloth Gemma 4 E2B IT QAT GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF/raw/db01ae3ceeca98487bf3569814f832f5023cd48c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings", "kv_sharing" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 35 text layers, seven full-attention layers, 28 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, num_kv_shared_layers 20, one KV head, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, vocab_size_per_layer_input 262144, hidden_size_per_layer_input 256, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E2B IT QAT unquantized metadata", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "kv_adapter", "per_layer_embeddings" ], "notes": "The base QAT unquantized config records the same checked memory-relevant Gemma 4 E2B architecture. Existing audited Google QAT GGUF and compressed-tensors profiles already use this config as the source for shared-KV and PLE behavior." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "Unsloth Gemma 4 E2B IT QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF/tree/db01ae3ceeca98487bf3569814f832f5023cd48c", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf is 2186184768 bytes, exactly matching API gguf.totalFileSize. Nearby main and sidecar files have different sizes: UD-Q4_K_XL is 2620368960 bytes, mmproj-BF16 is 986833728 bytes, mmproj-F16 is 985654080 bytes, mmproj-F32 is 1903027008 bytes, mtp-gemma-4-E2B-it.gguf is 59234176 bytes, and the MTP sidecar folder contains BF16, F16, Q4_0, and Q8_0 variants." }, { "label": "Unsloth Gemma 4 E2B IT QAT UD-Q2_K_XL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF/resolve/db01ae3ceeca98487bf3569814f832f5023cd48c/gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 47 metadata entries and 541 tensors. The selected file is 2.186184768 GB, with tensor payloads starting at byte 15813600. Tensor spans total 2.170371168 GB across 4628569635 logical elements: per_layer_token_embd.weight 1.321205760 GB, token_embd.weight 0.103809024 GB, blk.* tensors 0.717823072 GB, per_layer_model_proj/proj_norm tensors 0.027526144 GB, output_norm.weight 0.000006144 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/file overhead accounts for 0.015813600 GB. Swept tensor spans excluding the per-layer token lookup table total 0.849165408 GB across 2279759395 logical elements. Tensor spans split into Q4_0 1.716682752 GB, TQ2_0 0.395771904 GB, Q8_0 0.029245440 GB, F16 0.027525120 GB, and F32 0.001145952 GB. The header records gemma4.block_count 35, context_length 131072, attention.head_count 8, one KV head, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, shared_kv_layers 20, embedding_length_per_layer_input 256, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card metadata, immutable repo config, QAT unquantized config comparison, existing Gemma 4 E2B QAT profile comparison, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q2_K_XL artifact." }, "notes": "Use this profile for the Unsloth selected main UD-Q2_K_XL GGUF text artifact. Do not infer multimodal projector residency, audio/vision encoder traffic, or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--gemma-4-e4b-it-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-E4B-it-GGUF", "title": "Unsloth Gemma 4 E4B IT GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Gemma 4 E4B IT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it", "relation": "quantized", "source": "Hugging Face model card/API GGUF metadata, Google base config, Transformers Gemma 4 PLE documentation, and direct GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E4B-it. The selected GGUF header records the same Gemma 4 E4B text geometry as the Google config. The Unsloth repo does not ship config.json, so the immutable Google config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.51806929, "swept_params_b": 4.699497002, "auxiliary_resident_params_b": 2.818572288, "resident_weight_gb": 15.05309584, "swept_weight_gb": 9.400126784, "auxiliary_resident_weight_gb": 5.652969056, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-4-E4B-it-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges GGUF tensor spans in the selected main artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; separate mmproj and MTP GGUF files are not included unless explicitly loaded for another workload", "notes": "Gemma 4 E4B uses per-layer embeddings. Transformers documents a token-identity lookup from embed_tokens_per_layer and a context-aware per_layer_model_projection. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per_layer_proj_norm.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The Google config records global_head_dim 512 and attention_k_eq_v false, and the selected GGUF header contains separate attn_k and attn_v tensors for every layer." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj and MTP sidecars require separate workload profiles if loaded." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF because HF API gguf.totalFileSize matches gemma-4-E4B-it-BF16.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param records the BF16 tensor storage format." }, "evidence": [ { "label": "Unsloth Gemma 4 E4B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E4B-it-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit e1d90e5fb9f61d8dc71ef016580784a054e5c787 records base_model google/gemma-4-E4B-it, Apache-2.0 license, image-text-to-text pipeline, region:us, 630123 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 7518069290, and gguf.totalFileSize 15053095840." }, { "label": "Unsloth Gemma 4 E4B GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model google/gemma-4-E4B-it, and an Unsloth GGUF package with main text, mmproj, and MTP sidecar files." }, { "label": "Google Gemma 4 E4B IT config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings" ], "notes": "The immutable config records Gemma4ForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Transformers Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4", "source_type": "manual_review", "supports": [ "swept_parameter_scope", "auxiliary_scope", "per_layer_embeddings" ], "notes": "The Gemma 4 docs describe Per-Layer Embeddings as a token-identity lookup from embed_tokens_per_layer plus a context-aware per_layer_model_projection linear layer. This supports keeping per_layer_token_embd.weight resident-only while charging per_layer_model_proj.weight and block PLE projection tensors as swept matrix traffic." }, { "label": "Unsloth Gemma 4 E4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/tree/e1d90e5fb9f61d8dc71ef016580784a054e5c787", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-E4B-it-BF16.gguf is 15053095840 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files, mmproj-BF16/F16/F32.gguf, mtp-gemma-4-E4B-it.gguf, and MTP/* sidecar GGUFs have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Gemma 4 E4B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/resolve/e1d90e5fb9f61d8dc71ef016580784a054e5c787/gemma-4-E4B-it-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 52 metadata entries and 720 tensors. The linked file is 15.05309584 GB. Tensor spans sum to 15.03727136 GB: per_layer_token_embd.weight 5.637144576 GB / 2818572288 logical elements, token_embd.weight 1.34217728 GB, block tensors 8.002886976 GB, per_layer_model_proj/proj_norm tensors 0.055051264 GB, output_norm.weight 0.00001024 GB, and rope_freqs.weight 0.000001024 GB. Metadata/tokenizer/header/alignment bytes account for 0.01582448 GB. Swept tensor spans excluding the per-layer token lookup table total 9.400126784 GB / 4699497002 logical elements. The header records gemma4.block_count 42, context_length 131072, attention.head_count 8, attention.head_count_kv 2, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, embedding_length_per_layer_input 256, and separate attn_k and attn_v tensors for every layer. The selected main file has no mmproj, vision, audio, MTP, or output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, immutable Google config, Transformers Gemma 4 PLE documentation, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--gemma-4-e4b-it-qat-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gemma-4-E4B-it-qat-GGUF", "title": "Unsloth Gemma 4 E4B IT QAT GGUF UD-Q4_K_XL", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth UD-Q4_K_XL GGUF artifact of Gemma 4 E4B IT QAT.", "model_family": "gemma4-dense-multimodal-ple", "base_model_proof": { "base_model": "google/gemma-4-E4B-it-qat-q4_0-unquantized", "relation": "quantized", "source": "Hugging Face model card base_model metadata, Unsloth GGUF API metadata, repo config.json, base repo config.json, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of google/gemma-4-E4B-it-qat-q4_0-unquantized. Manual comparison found matching architecture, model_type, tied embedding, text layer, attention, sliding-window, PLE, and context fields between the packaged config and the Google QAT unquantized base config, except that the package config explicitly records torch_dtype bfloat16 while the base config omits the top-level field." }, "architecture": { "canonical_architecture_id": "gemma-4-e4b-qat", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.463013674, "swept_params_b": 4.644441386, "auxiliary_resident_params_b": 2.818572288, "resident_weight_gb": 4.21569376, "swept_weight_gb": 2.61442592, "auxiliary_resident_weight_gb": 1.60126784, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf", "swept_parameter_scope": "ordinary text decode charges GGUF tensor spans in the selected main artifact except per_layer_token_embd.weight; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "per_layer_token_embd.weight plus GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as full matrices per generated token; separate mmproj and MTP GGUF files are not included unless explicitly loaded for another workload", "notes": "Gemma 4 E4B uses per-layer embeddings. This profile treats per_layer_token_embd.weight as a resident lookup table rather than full swept matrix traffic, but includes per_layer_model_proj.weight, per_layer_proj_norm.weight, block tensors, token_embd.weight, output_norm.weight, and rope_freqs.weight in swept text-decode traffic." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 7, "kv_heads": 2, "head_dim": 512, "kv_scalar_multiplier": 2, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, and 41. The config records global_head_dim 512 and attention_k_eq_v false, and the selected GGUF header records 2 KV heads with 512 key/value length for global attention." }, { "kind": "sliding_window", "layers": 35, "kv_heads": 2, "head_dim": 256, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "Sliding-window layers reserve and read only up to the 512-token window recorded in the text config and GGUF metadata." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main UD-Q4_K_XL GGUF artifact after any multimodal prefill. The separate mmproj and MTP sidecars require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5648782039200643, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-0-qat-ple-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth UD-Q4_K_XL GGUF because the model card and generated usage examples explicitly load unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL. The HF API gguf.totalFileSize currently matches the smaller UD-Q2_K_XL sibling, so artifact selection is manual and documented rather than inferred from that API field." }, "evidence": [ { "label": "Unsloth Gemma 4 E4B QAT GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E4B-it-qat-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "api_artifact_mismatch" ], "notes": "The live HF API response at commit bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96 records base_model google/gemma-4-E4B-it-qat-q4_0-unquantized, Apache-2.0 license, any-to-any pipeline, endpoints_compatible, region:us, conversational, 236805 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 7463013674, and gguf.totalFileSize 3219530176. HEAD checks show that size matches the smaller UD-Q2_K_XL sibling, not the model-card-selected UD-Q4_K_XL file." }, { "label": "Unsloth Gemma 4 E4B QAT GGUF model card", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/raw/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact", "mtp_sidecar" ], "notes": "The card records Apache-2.0 licensing, base_model google/gemma-4-E4B-it-qat-q4_0-unquantized, QAT release context, and llama.cpp/Ollama/agent usage commands that select unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL. It also states the repo ships an MTP drafter sidecar at mtp-gemma-4-E4B-it.gguf, which is not included in this ordinary target-model profile." }, { "label": "Unsloth Gemma 4 E4B QAT GGUF config", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/raw/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings", "per_layer_embeddings", "kv_adapter" ], "notes": "The config records Gemma4ForConditionalGeneration, gemma4, bfloat16 source dtype, tie_word_embeddings true, 42 text layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max position embeddings, 262144 vocab_size_per_layer_input, resident vision config, and resident audio config." }, { "label": "Google Gemma 4 E4B IT QAT unquantized base config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json", "source_type": "config", "supports": [ "base_model_proof", "config_compatible", "model_family", "layers", "kv_adapter", "per_layer_embeddings" ], "notes": "The immutable QAT unquantized config matches the packaged config on the checked architecture fields: Gemma4ForConditionalGeneration, gemma4, tied embeddings, 42 layers, seven full-attention layers, 35 sliding-attention layers, 512-token sliding window, attention_k_eq_v false, 2 KV heads, 256 sliding head dimension, 512 global head dimension, 131072 max positions, and 262144 vocab_size_per_layer_input." }, { "label": "Unsloth Gemma 4 E4B QAT GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/tree/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "sidecar_scope", "api_artifact_mismatch" ], "notes": "HEAD checks found gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf is 4215693760 bytes, while gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf is 3219530176 bytes and matches API gguf.totalFileSize. Sidecars are separate: mmproj-BF16.gguf 991552320 bytes, mmproj-F16.gguf 990372672 bytes, mmproj-F32.gguf 1912464192 bytes, mtp-gemma-4-E4B-it.gguf 59676544 bytes, and MTP/gemma-4-E4B-it-BF16-MTP.gguf 171784064 bytes." }, { "label": "Unsloth Gemma 4 E4B QAT UD-Q4_K_XL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/resolve/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 47 metadata entries and 666 tensors. The selected linked file is 4.215693760 GB, with tensor payloads starting at byte 15820928. Tensor spans total 4.199872832 GB across 7463013674 logical elements. per_layer_token_embd.weight is 1.585446912 GB / 2818572288 logical elements and is resident-only for this ordinary text-decode profile. Swept tensor spans excluding that lookup table total 2.614425920 GB / 4644441386 logical elements. token_embd.weight is 0.377487360 GB and no output.weight tensor is stored, so token_embd.weight remains swept as tied output-projection traffic. The header records general.name Gemma-4 E4B IT (smart Q4_0, QAT-lossless), general.quantized_by unsloth, gemma4.block_count 42, context_length 131072, attention.head_count 8, attention.head_count_kv 2, key/value length 512 for global attention, key/value length 256 for sliding-window attention, sliding_window 512, and embedding_length_per_layer_input 256." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned model card, pinned package config, pinned Google QAT unquantized base config, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth main UD-Q4_K_XL target-model text artifact. Do not infer multimodal projector, MTP drafter, or smaller UD-Q2_K_XL residency unless those separate GGUF files are explicitly selected by the workload." }, { "id": "unsloth--glm-4-7-flash-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/GLM-4.7-Flash-GGUF", "title": "Unsloth GLM 4.7 Flash GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected GLM-4.7-Flash IQ4_NL GGUF artifact.", "model_family": "glm4-moe-lite-gguf", "base_model_proof": { "base_model": "zai-org/GLM-4.7-Flash", "relation": "quantized", "source": "Unsloth model card, live HF GGUF metadata, pinned Z.ai base config, direct GGUF header range read, and current llama.cpp DeepSeek2/cache source review", "config_compatible": true, "notes": "The Unsloth repo identifies zai-org/GLM-4.7-Flash as the base model. The selected GGUF header uses llama.cpp's deepseek2 architecture label, but records the audited GLM-4.7-Flash geometry: 47 ordinary blocks, first dense block 1, 64 routed experts, 4 experts per token, one shared expert, 202752 context, q_lora_rank 768, kv_lora_rank 512, and MLA key/value dimensions." }, "architecture": { "canonical_architecture_id": "glm-4-7-flash", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.165033824, "main_resident_weight_gb": 16.977136128, "auxiliary_resident_weight_gb": 0.187897696, "fixed_weight_gb": 1.349159424, "routed_expert_weight_gb": 0.244187136, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected GLM-4.7-Flash-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary llama.cpp text decode through blk.0 through blk.46 plus output.weight and output_norm.weight from the selected IQ4_NL GGUF artifact, excluding input embedding lookup", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept as full matrices for each ordinary generated token; the selected main file has no blk.47, MTP/NextN, DSA indexer, or multimodal sidecar tensors", "shared_expert_notes": "The GGUF metadata records expert_shared_count 1. Shared expert branch tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 30B/3B model-card parameters. The selected API artifact is GLM-4.7-Flash-IQ4_NL.gguf. Routed expert tensors are byte-uniform across the 64 expert indexes in ordinary MoE layers 1-46." }, "kv_adapter": { "kind": "full_context", "layers": 47, "kv_heads": 1, "head_dim": 576, "kv_scalar_multiplier": 1, "notes": "The selected GGUF header records deepseek2.attention.head_count_kv 1 and attention.key_length 576. Current llama.cpp allocates no V cache for MLA models, so the cache is compact FP16 K-only rather than expanded full K/V. This gives 47 * 1 * 576 * 2 bytes per cached token, or 0.054144 GB per 1K context tokens." }, "notes": "This profile models ordinary llama.cpp text decode for the selected IQ4_NL GGUF artifact. It intentionally does not reuse the audited BF16 Transformers profile's expanded K/V cache or the MLX artifact's safetensors byte layout, because the GGUF runtime path and selected quantized tensor spans differ." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.573249440923763, "kv_store_format": "llama.cpp-compact-mla-k-only-fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "llama.cpp-compact-mla-k-only-fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-deepseek2-mla-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF linked bytes for residency and exact tensor spans for swept weight traffic. GGUF loader overhead, tokenizer work, kernels, routing compute, sampling, scheduler behavior, and speculative decoding are outside Bounds Engine v1.", "notes": "The selected artifact stores most tensors as IQ4_NL plus Q4_K token embeddings, Q6_K output weights, Q8_0 MLA side tensors, Q5_K/Q6_K/F32 side tensors, and GGUF metadata/header bytes. The profile uses FP16 KV because the selected GGUF does not declare a required quantized KV-cache format." }, "evidence": [ { "label": "Unsloth GLM-4.7-Flash GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/GLM-4.7-Flash-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "pipeline", "selected_artifact", "total_params_b", "max_context_tokens", "license" ], "notes": "At commit 0d32489ecb9db6d2a4fc93bd27ef01519f95474d, the live API records a public non-gated MIT text-generation GGUF repo with transformers, unsloth, en/zh, endpoints_compatible, deploy:azure, region:us, conversational, imatrix, and base_model:quantized:zai-org/GLM-4.7-Flash tags. Current downloads are 94041, below the local catalog's older qualifying snapshot. The API GGUF block records architecture deepseek2, context_length 202752, total 29943393920, and totalFileSize 17165033824." }, { "label": "Unsloth GLM-4.7-Flash GGUF model card", "url": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/raw/0d32489ecb9db6d2a4fc93bd27ef01519f95474d/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "pipeline", "language", "serving" ], "notes": "The card metadata records base_model zai-org/GLM-4.7-Flash, language en/zh, MIT license, transformers library, text-generation pipeline, and unsloth tags. The visible card says GLM-4.7-Flash is a 30B-A3B MoE model, notes Unsloth Dynamic 2.0 GGUF, documents a llama.cpp output-quality update, and includes vLLM/SGLang/Transformers serving guidance. It does not pin a different GGUF suffix for the repo-level selected artifact." }, { "label": "Z.ai GLM-4.7-Flash base config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned base config records Glm4MoeLiteForCausalLM, glm4_moe_lite, BF16 dtype, 47 hidden layers, num_nextn_predict_layers 1, first_k_dense_replace 1, hidden size 2048, intermediate size 10240, moe_intermediate_size 1536, 20 attention heads, 20 KV heads in the Transformers config, q_lora_rank 768, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, v_head_dim 256, 64 routed experts, 4 experts per token, one shared expert, untied embeddings, vocab size 154880, and max_position_embeddings 202752." }, { "label": "Unsloth GLM-4.7-Flash GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/tree/0d32489ecb9db6d2a4fc93bd27ef01519f95474d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found the API-selected IQ4_NL file at 17.165033824 GB, matching API gguf.totalFileSize exactly. Nearby siblings include IQ4_XS 16.271360352 GB, MXFP4_MOE 16.968499296 GB, Q4_K_M 18.312339808 GB, Q8_0 31.842799968 GB, UD-IQ1_M 9.808507232 GB, UD-Q4_K_XL 17.520169312 GB, imatrix_unsloth.gguf_file 0.072447904 GB, and the two BF16 split shards totaling 59.908837696 GB." }, { "label": "Unsloth GLM-4.7-Flash IQ4_NL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/resolve/0d32489ecb9db6d2a4fc93bd27ef01519f95474d/GLM-4.7-Flash-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format" ], "notes": "A 96MB range-read of the selected GGUF v3 header found 60 metadata entries and 844 tensors. The linked file is 17.165033824 GB. Tensor spans sum to 17.155557888 GB; metadata/tokenizer/header/file overhead accounts for 0.009475936 GB. Tensor spans split into IQ4_NL 16.086122496 GB, Q6_K 0.378900480 GB, Q8_0 0.287997952 GB, Q5_K 0.198967296 GB, Q4_K 0.178421760 GB, and F32 0.025147904 GB. token_embd.weight is 0.178421760 GB and resident-only; output.weight plus output_norm.weight total 0.260206592 GB and are swept. Ordinary text traffic, excluding token embedding, is 16.977136128 GB. Routed expert tensors in ordinary MoE layers 1-46 sum to 15.627976704 GB, or 0.244187136 GB per uniform expert index. Fixed ordinary text traffic, including layer 0 dense MLP, attention, routers, shared experts, norms, output.weight, and output_norm.weight, sums to 1.349159424 GB. The selected main file contains no blk.47, MTP/NextN, DSA indexer, vision, visual, image, or mmproj tensors. The GGUF metadata records deepseek2 block_count 47, context_length 202752, attention.head_count 20, attention.head_count_kv 1, attention.key_length 576, attention.value_length 512, attention.kv_lora_rank 512, attention.q_lora_rank 768, attention.key_length_mla 256, attention.value_length_mla 256, expert_count 64, expert_used_count 4, expert_shared_count 1, and leading_dense_block_count 1." }, { "label": "llama.cpp DeepSeek2 and cache source review", "url": "https://github.com/ggml-org/llama.cpp/blob/cb295bf59663cd3577389315636772f4060bd1f5/src/models/deepseek2.cpp", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual source review at llama.cpp commit cb295bf59663cd3577389315636772f4060bd1f5 found the DeepSeek2 loader creates MLA tensors from q_lora_rank and kv_lora_rank metadata, creates split attn_k_b and attn_v_b tensors for MLA, creates routed expert tensors from ffn_down_exps, ffn_gate_exps, and ffn_up_exps, and treats ffn_exp_probs_b as optional. The DeepSeek2 graph builds Kcur by concatenating kv_cmpr with k_pe and passes Vcur as kv_cmpr. llama-kv-cache.cpp allocates has_v = !is_mla, so MLA models use a compact K-only cache." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, the pinned Unsloth model card, pinned Z.ai base config, linked-object HEAD checks, a direct GGUF header/tensor-index range read of the API-selected IQ4_NL artifact, and current llama.cpp DeepSeek2/cache source review." }, "notes": "Use this profile for the API-selected IQ4_NL GGUF memory-side bound. Do not silently substitute MXFP4_MOE, Q4_K_M, UD dynamic, or BF16 split siblings; those require separate selected-artifact profiles." }, { "id": "unsloth--glm-5-2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/GLM-5.2-GGUF", "title": "Unsloth GLM-5.2 GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected GLM-5.2 BF16 GGUF split artifact.", "model_family": "glm-5-2-moe-dsa-gguf", "base_model_proof": { "base_model": "zai-org/GLM-5.2", "relation": "quantized", "source": "Unsloth model card, live HF GGUF metadata, pinned Z.ai base config, direct GGUF header range reads, and current llama.cpp GLM_DSA source review", "config_compatible": true, "notes": "The Unsloth repo identifies zai-org/GLM-5.2 as the base model and the live GGUF metadata records the same 1M-context GLM DSA architecture. Manual comparison against the pinned base config found matching ordinary text geometry: 78 hidden layers, one next-token-prediction auxiliary layer, first three dense layers, hidden size 6144, feed-forward size 12288, MoE feed-forward size 2048, 64 attention heads, q_lora_rank 2048, kv_lora_rank 512, 256 routed experts, 8 routed experts per token, one shared expert, untied embeddings, vocabulary size 154880, and max context 1048576." }, "architecture": { "canonical_architecture_id": "glm-5-2-llamacpp-gguf", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1507.988023008, "main_resident_weight_gb": 1484.695550976, "auxiliary_resident_weight_gb": 23.292472032, "fixed_weight_gb": 35.144088576, "routed_expert_weight_gb": 5.6623104, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected 33-shard BF16 GGUF linked file size from HF API and direct linked-object HEAD checks", "traffic_scope": "current llama.cpp ordinary text decode through layers 0-77 using the GLM_DSA alias of the DeepSeek2 graph; excludes input embedding lookup, layer 78 MTP/NextN tensors, GGUF overhead, and optional DSA indexer tensors that are loaded as not-required but not used by the current graph", "auxiliary_scope": "GGUF metadata/header/alignment, token_embd.weight, layer 78 MTP/NextN tensors, and current llama.cpp-unused DSA indexer tensors are resident in the selected artifact but are not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert branch tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "A direct GGUF tensor-index audit split linked bytes into 1449.551462400 GB routed expert tensors, 33.240898560 GB used non-expert block tensors, 1.903190016 GB output/output_norm tensors, 1.903165440 GB token embedding, 19.909447680 GB layer-78 MTP/NextN tensors, 1.470314496 GB current llama.cpp-unused indexer tensors, and 0.009544416 GB GGUF metadata/header/alignment overhead." }, "kv_adapter": { "kind": "full_context", "layers": 78, "kv_heads": 1, "head_dim": 576, "kv_scalar_multiplier": 1, "notes": "Current llama.cpp routes GLM_DSA through the standard llama_kv_cache path, not the DSA two-cache wrapper used by DeepSeek DSA variants. For MLA hparams, llama_kv_cache stores K only and omits V. The selected GGUF header records attention.head_count_kv 1 and attention.key_length 576, so the compact FP16 cache is 78 * 1 * 576 scalars per token." }, "notes": "This profile models ordinary llama.cpp text decode for the selected BF16 split GGUF artifact. It intentionally does not reuse the audited Transformers GLM-5.2-FP8 expanded K/V plus IndexShare indexer profile because the current GGUF runtime path has different cache allocation and does not execute the optional indexer tensors." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000344551463002, "kv_store_format": "llama.cpp-compact-mla-k-only-fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "llama.cpp-compact-mla-k-only-fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-glm-dsa-deepseek2-graph-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF linked bytes for residency and exact tensor spans for swept weight traffic. GGUF loader overhead, tokenizer work, kernels, routing compute, sampling, scheduler behavior, and speculative decoding are outside Bounds Engine v1.", "notes": "The selected split artifact stores almost all tensors as BF16 plus small F32 bias/norm tensors. The profile uses FP16 KV because GGUF does not declare a required quantized KV-cache format." }, "evidence": [ { "label": "Unsloth GLM-5.2 GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/GLM-5.2-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "pipeline", "selected_artifact", "total_params_b", "max_context_tokens", "license" ], "notes": "At commit abc55e72527792c6e77069c99b4cb7de16fa9f23, the live API records a public non-gated MIT text-generation GGUF repo with glm_moe_dsa, unsloth, en/zh, endpoints_compatible, region:us, conversational, and base_model:quantized:zai-org/GLM-5.2 tags. Current downloads are 392857. The API GGUF block records architecture glm-dsa, context_length 1048576, total 753864139008, and totalFileSize 1507988023008." }, { "label": "Unsloth GLM-5.2 GGUF model card", "url": "https://huggingface.co/unsloth/GLM-5.2-GGUF/raw/abc55e72527792c6e77069c99b4cb7de16fa9f23/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "pipeline", "language", "serving" ], "notes": "The card metadata records base_model zai-org/GLM-5.2, language en/zh, MIT license, text-generation pipeline, and glm_moe_dsa/unsloth tags. The bundled upstream card describes GLM-5.2 as a 1M-context GLM MoE DSA model with IndexShare and MTP/speculative-decoding support." }, { "label": "Z.ai GLM-5.2 base config", "url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned base config records GlmMoeDsaForCausalLM, glm_moe_dsa, bfloat16 dtype, 78 hidden layers, num_nextn_predict_layers 1, first_k_dense_replace 3, hidden size 6144, intermediate size 12288, moe_intermediate_size 2048, 64 attention heads, 64 KV heads in the Transformers config, q_lora_rank 2048, kv_lora_rank 512, qk_head_dim 256, qk_nope_head_dim 192, qk_rope_head_dim 64, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, one shared expert, untied embeddings, vocab size 154880, max_position_embeddings 1048576, and an indexer_types list with 21 full and 57 shared entries." }, { "label": "Unsloth GLM-5.2 BF16 GGUF split header audit", "url": "https://huggingface.co/unsloth/GLM-5.2-GGUF/tree/abc55e72527792c6e77069c99b4cb7de16fa9f23/BF16", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format" ], "notes": "Direct HEAD checks and range-reads of all 33 selected BF16 GGUF shard headers matched API gguf.totalFileSize at 1507.988023008 GB linked bytes. Tensor spans sum to 1507.978478592 GB over 753.864139008B logical elements, with 0.009544416 GB metadata/header/alignment overhead. Tensor spans split into 1449.551462400 GB routed expert tensors, 33.240898560 GB current llama.cpp-used non-expert block tensors, 1.903190016 GB output/output_norm tensors, 1.903165440 GB token_embd.weight, 19.909447680 GB layer-78 MTP/NextN tensors, and 1.470314496 GB optional DSA indexer tensors. The GGUF metadata records glm-dsa, block_count 79, context_length 1048576, attention.head_count 64, attention.head_count_kv 1, attention.key_length 576, attention.value_length 512, q_lora_rank 2048, kv_lora_rank 512, attention.key_length_mla 256, attention.value_length_mla 256, expert_count 256, expert_used_count 8, expert_shared_count 1, expert_feed_forward_length 2048, indexer.head_count 32, indexer.key_length 128, and indexer.top_k 2048." }, { "label": "llama.cpp GLM_DSA and cache source review", "url": "https://github.com/ggml-org/llama.cpp/blob/cb295bf59663cd3577389315636772f4060bd1f5/src/models/glm-dsa.cpp", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual source review at llama.cpp commit cb295bf59663cd3577389315636772f4060bd1f5 found GLM_DSA requires MLA, loads token_embd, output/output_norm, ordinary layers, optional DSA indexer tensors as TENSOR_NOT_REQUIRED, and layer-78 NextN/MTP tensors as skipped/not-required. models.h aliases llama_model_glm_dsa::graph to llama_model_deepseek2::graph. llama-model.cpp only instantiates llama_kv_cache_dsa for DeepSeek DSA variants, while GLM_DSA falls through to the standard llama_kv_cache path. llama-kv-cache.cpp omits V cache for hparams.is_mla(), leaving a compact K-only cache." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, the pinned Unsloth model card, pinned Z.ai base config, linked-object HEAD checks, direct GGUF header/tensor-index range reads across all selected BF16 split shards, and current llama.cpp GLM_DSA/cache source review." }, "notes": "Use this profile for the API-selected BF16 split GGUF memory-side bound. Do not silently substitute smaller Q8_0, Q4_K_M, dynamic GGUF, or other quantized siblings; those require separate selected-artifact profiles." }, { "id": "unsloth--gpt-oss-120b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gpt-oss-120b-GGUF", "title": "Unsloth gpt-oss 120B GGUF F16", "summary": "Audited memory-side bounds profile for the API-selected Unsloth F16 GGUF artifact of gpt-oss-120b.", "model_family": "gpt-oss-moe", "base_model_proof": { "base_model": "openai/gpt-oss-120b", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, HF API GGUF metadata, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of openai/gpt-oss-120b. The served config matches the pinned OpenAI base config on checked architecture fields, and the selected GGUF header records the same gpt-oss geometry: 36 layers, 128 local experts, 4 experts per token, alternating sliding/full attention, 128-token sliding window, 8 KV heads, 64 key/value dimensions, and 131072 context length." }, "architecture": { "canonical_architecture_id": "gpt-oss-120b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 65.369017728, "main_resident_weight_gb": 64.197727488, "auxiliary_resident_weight_gb": 1.17129024, "fixed_weight_gb": 3.124401408, "routed_expert_weight_gb": 0.47713536, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected gpt-oss-120b-F16.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.35 non-routed tensors, routers, expert biases, and expected-distinct routed expert weight groups from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "The selected F16 GGUF file stores routed expert weights as MXFP4 tensors while storing attention, embeddings, lm_head, norms, routers, and expert biases as F16/F32 tensors. Routed expert weights and expert biases are byte-uniform across the 128 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 18, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "Odd-numbered layers in the config use full_attention." }, { "kind": "sliding_window", "layers": 18, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "kv_scalar_multiplier": 2, "notes": "Even-numbered layers in the config use sliding_attention with a 128-token window." } ], "notes": "The served config and selected GGUF metadata alternate sliding_attention and full_attention across 36 layers." }, "notes": "This profile models ordinary text decode for the API-selected main F16 GGUF artifact. Other GGUF quantizations in the same repo have different resident and traffic bytes and require separate workload selection if used." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5595265735892001, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-gpt-oss-f16-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and speculative behavior are outside Bounds Engine v1.", "notes": "The HF API selected artifact is gpt-oss-120b-F16.gguf. The selected GGUF contains MXFP4 expert weights and F16/F32 side tensors; KV is charged as FP16 for llama.cpp-style GGUF serving." }, "evidence": [ { "label": "Unsloth gpt-oss 120B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gpt-oss-120b-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit ff1a82da6ad466e32284fa3d2b86694db3204789 records a public Apache-2.0 text-generation GGUF repo with base_model openai/gpt-oss-120b, 184,379 downloads, region:us, GGUF architecture gpt-oss, 131072 context length, gguf.total 116829156672, and gguf.totalFileSize 65369017728. The API totalFileSize matches gpt-oss-120b-F16.gguf, so this profile targets that artifact." }, { "label": "Unsloth gpt-oss 120B GGUF model card", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "serving" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model openai/gpt-oss-120b, llama.cpp support, and multiple GGUF quantizations. It identifies the F16 quant as the original-precision gpt-oss package and provides local serving guidance." }, { "label": "Unsloth gpt-oss 120B GGUF served config", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF/raw/ff1a82da6ad466e32284fa3d2b86694db3204789/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "attention_pattern", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The served config records GptOssForCausalLM, gpt_oss, BF16, 36 layers, hidden size 2880, 64 attention heads, 8 KV heads, 64 head dimension, 128 local experts, 4 experts per token, alternating sliding_attention and full_attention layers, 128-token sliding window, untied embeddings, vocab size 201088, and 131072 max position embeddings. Manual comparison against openai/gpt-oss-120b at commit b5c939de8f754692c1647ca79fbf85e8c1e70f8a found no differences in checked architecture fields." }, { "label": "Unsloth gpt-oss 120B F16 GGUF linked object and range-read tensor index", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF/resolve/ff1a82da6ad466e32284fa3d2b86694db3204789/gpt-oss-120b-F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "attention_pattern", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 37 metadata entries and 687 tensors. The linked file is 65.369017728 GB. Tensor spans sum to 65.355994368 GB; metadata/tokenizer/header/file overhead accounts for 0.013023360 GB. Tensor spans split into MXFP4 60.914073600 GB, F16 4.227563520 GB, and F32 0.214357248 GB. token_embd.weight is 1.158266880 GB and resident-only; output.weight is 1.158266880 GB and swept. Routed expert weights plus expert biases sum to 61.073326080 GB, or 0.477135360 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 3.124401408 GB. The header records gpt-oss block_count 36, context_length 131072, expert_count 128, expert_used_count 4, alternating full/sliding attention, 8 KV heads, 64 key/value length, and 128-token sliding window." }, { "label": "Unsloth gpt-oss 120B GGUF sibling HEAD checks", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF/tree/ff1a82da6ad466e32284fa3d2b86694db3204789", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found gpt-oss-120b-F16.gguf is 65.369017728 GB, exactly matching API gguf.totalFileSize. Split quantization directory sums differ: Q2_K 62.570256992 GB, Q2_K_L 62.859823712 GB, Q3_K_M 62.626843232 GB, Q3_K_S 62.563621472 GB, Q4_0 62.620023392 GB, Q4_1 62.715938912 GB, Q4_K_M 62.768723552 GB, Q4_K_S 62.759138912 GB, Q5_K_M 62.889522272 GB, Q5_K_S 62.881227872 GB, Q6_K 63.284496960 GB, Q8_0 63.387347520 GB, UD-Q4_K_XL 63.016311392 GB, UD-Q6_K_XL 63.284496960 GB, and UD-Q8_K_XL 64.473222720 GB. The API-selected artifact is F16." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served config, OpenAI base config comparison, selected linked file size, all sibling linked file HEAD checks, and a direct GGUF header/tensor-index range read of the API-selected F16 artifact." }, "notes": "Use this profile for the API-selected Unsloth F16 GGUF main artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations unless the workload explicitly selects those files." }, { "id": "unsloth--gpt-oss-20b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/gpt-oss-20b-GGUF", "title": "Unsloth gpt-oss 20B GGUF F16", "summary": "Audited memory-side bounds profile for the API-selected Unsloth F16 GGUF artifact of gpt-oss-20b.", "model_family": "gpt-oss-moe", "base_model_proof": { "base_model": "openai/gpt-oss-20b", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served config comparison, HF API GGUF metadata, selected linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of openai/gpt-oss-20b. The served config matches the pinned OpenAI base config on checked architecture fields, and the selected GGUF header records the same gpt-oss geometry: 24 layers, 32 local experts, 4 experts per token, alternating sliding/full attention, 128-token sliding window, 8 KV heads, 64 key/value dimensions, and 131072 context length." }, "architecture": { "canonical_architecture_id": "gpt-oss-20b", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 13.792639168, "main_resident_weight_gb": 12.621363456, "auxiliary_resident_weight_gb": 1.171275712, "fixed_weight_gb": 2.442475776, "routed_expert_weight_gb": 0.31809024, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected gpt-oss-20b-F16.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.23 non-routed tensors, routers, expert biases, and expected-distinct routed expert weight groups from the selected F16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "The selected F16 GGUF file stores routed expert weights as MXFP4 tensors while storing attention, embeddings, lm_head, norms, routers, and expert biases as F16/F32 tensors. Routed expert weights and expert biases are byte-uniform across the 32 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "Odd-numbered layers in the config use full_attention." }, { "kind": "sliding_window", "layers": 12, "kv_heads": 8, "head_dim": 64, "window_tokens": 128, "kv_scalar_multiplier": 2, "notes": "Even-numbered layers in the config use sliding_attention with a 128-token window." } ], "notes": "The served config and selected GGUF metadata alternate sliding_attention and full_attention across 24 layers." }, "notes": "This profile models ordinary text decode for the API-selected main F16 GGUF artifact. Other GGUF quantizations in the same repo have different resident and traffic bytes and require separate workload selection if used." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6594692468412451, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-gpt-oss-f16-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and speculative behavior are outside Bounds Engine v1.", "notes": "The HF API selected artifact is gpt-oss-20b-F16.gguf. The selected GGUF contains MXFP4 expert weights and F16/F32 side tensors; KV is charged as FP16 for llama.cpp-style GGUF serving." }, "evidence": [ { "label": "Unsloth gpt-oss 20B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/gpt-oss-20b-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4 records a public Apache-2.0 text-generation GGUF repo with base_model openai/gpt-oss-20b, 405984 downloads, region:us, GGUF architecture gpt-oss, 131072 context length, gguf.total 20914757184, and gguf.totalFileSize 13792639168. The API totalFileSize matches gpt-oss-20b-F16.gguf, so this profile targets that artifact." }, { "label": "Unsloth gpt-oss 20B GGUF model card", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "serving" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model openai/gpt-oss-20b, llama.cpp support, and multiple GGUF quantizations. It notes native MXFP4 quantization for gpt-oss MoE layers and provides local serving examples." }, { "label": "Unsloth gpt-oss 20B GGUF served config", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF/raw/d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "attention_pattern", "sliding_window", "routed_experts", "routed_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The served config records GptOssForCausalLM, gpt_oss, BF16, 24 layers, hidden size 2880, 64 attention heads, 8 KV heads, 64 head dimension, 32 local experts, 4 experts per token, alternating sliding_attention and full_attention layers, 128-token sliding window, untied embeddings, vocab size 201088, and 131072 max position embeddings. Manual comparison against openai/gpt-oss-20b found no differences in checked architecture fields." }, { "label": "Unsloth gpt-oss 20B F16 GGUF linked object and range-read tensor index", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF/resolve/d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4/gpt-oss-20b-F16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "attention_pattern", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 37 metadata entries and 459 tensors. The linked file is 13.792639168 GB. Tensor spans sum to 13.779630336 GB; metadata/tokenizer/header/file overhead accounts for 0.013008832 GB. Tensor spans split into MXFP4 10.152345600 GB, F16 3.590553600 GB, and F32 0.036731136 GB. token_embd.weight is 1.158266880 GB and resident-only; output.weight is 1.158266880 GB and swept. Routed expert weights plus expert biases sum to 10.178887680 GB, or 0.318090240 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 2.442475776 GB." }, { "label": "Unsloth gpt-oss 20B GGUF sibling HEAD checks", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF/tree/d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found gpt-oss-20b-F16.gguf 13.792639168 GB, Q2_K 11.468317888 GB, Q2_K_L 11.757884608 GB, Q3_K_M 11.506103488 GB, Q3_K_S 11.463894208 GB, Q4_0 11.501495488 GB, Q4_1 11.577504448 GB, Q4_K_M 11.624759488 GB, Q4_K_S 11.618492608 GB, Q5_K_M 11.717357248 GB, Q5_K_S 11.711827648 GB, Q6_K 12.041000128 GB, Q8_0 12.109567168 GB, UD-Q4_K_XL 11.872347328 GB, UD-Q6_K_XL 12.041000128 GB, and UD-Q8_K_XL 13.195442368 GB. The API-selected artifact is F16." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, served config, OpenAI base config comparison, selected linked file size, and a direct GGUF header/tensor-index range read of the API-selected F16 artifact." }, "notes": "Use this profile for the API-selected Unsloth F16 GGUF main artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations unless the workload explicitly selects those files." }, { "id": "unsloth--kimi-k2-7-code-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Kimi-K2.7-Code-GGUF", "title": "Unsloth Kimi K2.7 Code GGUF UD-IQ1_M Split", "summary": "Audited memory-side ordinary text-decode bounds profile for the API-selected eight-part UD-IQ1_M split GGUF artifact of Kimi K2.7 Code.", "model_family": "kimi-k2-moe", "base_model_proof": { "base_model": "moonshotai/Kimi-K2.7-Code", "relation": "quantized", "source": "Hugging Face model card/API metadata, repo config.json, audited Moonshot Kimi K2.7 Code profile, selected split-GGUF header metadata, and selected linked-object size checks", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a quantized GGUF derivative of moonshotai/Kimi-K2.7-Code. The repo config, base model card, and selected GGUF headers match the audited Kimi K2.7 Code layer count, MLA ranks, expert counts, routing shape, and context length." }, "architecture": { "canonical_architecture_id": "kimi-k2-7-code", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 303.909170464, "main_resident_weight_gb": 303.241582592, "auxiliary_resident_weight_gb": 0.667587872, "fixed_weight_gb": 8.524617728, "routed_expert_weight_gb": 0.767492096, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected eight-part UD-IQ1_M split GGUF linked-file size and API GGUF logical tensor parameters", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, attention tensors, shared/dense FFN tensors, router tensors, and expected distinct routed expert tensor spans from the selected UD-IQ1_M split; token_embd.weight is resident-only input embedding traffic", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, headers, and file overhead are resident in the selected split artifact but not swept for ordinary text decode; mmproj sidecar GGUF files are excluded unless explicitly loaded for an image/video workload", "shared_expert_notes": "The selected GGUF header records expert_shared_count 1. The dense layer-0 MLP and shared-expert tensors are charged in fixed_weight_gb because they are always-on traffic.", "notes": "HF API gguf.total is 1026.408232448B logical tensor parameters and gguf.totalFileSize selects the eight-part UD-IQ1_M split. Range-reads of all eight GGUF v3 shard headers found 1096 tensors. Linked split files total 303.909170464 GB. Tensor spans total 303.902185472 GB, while metadata/header/tokenizer/alignment overhead accounts for 0.006984992 GB. Routed expert tensors total 294.716964864 GB across layers 1-60 and 384 expert indexes, or 0.767492096 GB per expert index." }, "kv_adapter": { "kind": "compressed_state", "alloc_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Compact MLA cache coefficient: 61 layers * (kv_lora_rank 512 + qk_rope_head_dim 64) * 2 FP16 bytes * 1000 / 1e9." }, "read_formula": { "kind": "linear_context_ratio", "gb_per_1k_context_tokens": 0.070272, "notes": "Decode reads the same compact MLA latent plus RoPE cache bytes per active context token in this v1 memory-traffic approximation." }, "notes": "The selected GGUF header records deepseek2 compact MLA geometry: one KV head, key length 576, value length 512, kv_lora_rank 512, and rope dimension 64. This profile follows the compact MLA state policy used by the audited DeepSeek split-GGUF profile rather than the expanded Hugging Face custom-code cache path." }, "notes": "This profile models ordinary text decode for the API-selected UD-IQ1_M split GGUF artifact. It does not substitute the Q4/Q8 split families or include the separate mmproj sidecar files." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.29608995802690685, "kv_store_format": "llama_cpp_compact_mla_fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "llama_cpp_compact_mla_fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-iq1-m-kimi-k2-7-compact-mla-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected split-GGUF tensor spans for weight traffic and selected linked-file bytes for residency. GGUF split loading overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and non-selected GGUF variants are outside Bounds Engine v1.", "notes": "The API-selected artifact is the eight-part UD-IQ1_M split because the linked-file sum exactly matches HF API gguf.totalFileSize. Default llama.cpp-style GGUF KV/state is modeled as two-byte compact MLA state." }, "evidence": [ { "label": "Unsloth Kimi K2.7 Code GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Kimi-K2.7-Code-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 46352ca8dc32aa60f9754f5bd3fb778deeb9e430, the live API records a public non-gated GGUF repo with base_model moonshotai/Kimi-K2.7-Code, base_model_relation quantized, modified-MIT/other license metadata, region:us, imatrix metadata, 291513 downloads, GGUF architecture deepseek2, context length 262144, gguf.total 1026408232448, and gguf.totalFileSize 303909170464." }, { "label": "Unsloth Kimi K2.7 Code GGUF model card", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/raw/46352ca8dc32aa60f9754f5bd3fb778deeb9e430/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "runtime_format" ], "notes": "The card metadata records base_model moonshotai/Kimi-K2.7-Code, modified-MIT licensing, and GGUF packaging. The model summary table records 1T total parameters, 32B activated parameters, 61 layers, 384 experts, 8 selected experts per token, 1 shared expert, 160K vocabulary, 256K context, MLA attention, and MoonViT vision encoder. The card says UD-Q8_K_XL is the full-precision lossless variant and that lower-bit GGUF variants are separate artifacts." }, { "label": "Unsloth Kimi K2.7 Code GGUF config", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/raw/46352ca8dc32aa60f9754f5bd3fb778deeb9e430/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "multimodal" ], "notes": "The config records KimiK25ForConditionalGeneration wrapping DeepseekV3ForCausalLM, bfloat16 text dtype, 61 hidden layers, first_k_dense_replace 1, hidden size 7168, intermediate size 18432, max position embeddings 262144, 64 attention heads, 64 key/value heads, kv_lora_rank 512, q_lora_rank 1536, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, 384 routed experts, 8 experts per token, 1 shared expert, and compressed-tensors INT4 metadata inherited from the source package." }, { "label": "Unsloth Kimi K2.7 Code UD-IQ1_M split linked-object size checks", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/tree/46352ca8dc32aa60f9754f5bd3fb778deeb9e430/UD-IQ1_M", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_resident_weight_gb" ], "notes": "Exact HEAD checks found the eight UD-IQ1_M split shards total 303.909170464 GB, exactly matching API gguf.totalFileSize. The first metadata shard is 0.006913184 GB; tensor shards are 49.638722336, 48.368994848, 48.368994848, 49.246454080, 49.253143328, 49.885925408, and 9.140022432 GB. The separate mmproj sidecars are 0.953927072 GB BF16, 0.952572320 GB F16, and 1.884595616 GB F32 and are not included in this ordinary text profile." }, { "label": "Unsloth Kimi K2.7 Code UD-IQ1_M split GGUF header audit", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/46352ca8dc32aa60f9754f5bd3fb778deeb9e430/UD-IQ1_M/Kimi-K2.7-Code-UD-IQ1_M-00001-of-00008.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_adapter", "weight_format" ], "notes": "Range-reads of the selected UD-IQ1_M split GGUF v3 headers found split.count 8 and split.tensors.count 1096. Header tensor spans across all eight shards sum to 303.902185472 GB: routed expert tensors 294.716964864 GB, non-expert tensor spans 9.185220608 GB, output.weight 0.660602880 GB, token_embd.weight 0.660602880 GB, output_norm.weight 0.000028672 GB, and blk.* tensors 302.580951040 GB. token_embd.weight is resident-only because the selected split stores a separate output.weight. Fixed ordinary text traffic excluding token_embd.weight is 8.524617728 GB. Metadata/header/tokenizer/alignment overhead across the split accounts for 0.006984992 GB. Stored tensor spans split into IQ3_XXS 120.846286848 GB, IQ2_XXS 87.199580160 GB, IQ1_M 73.987522560 GB, Q4_0 12.683575296 GB, Q5_K 4.248961024 GB, Q8_0 2.050174976 GB, Q4_K 1.321205760 GB, Q6_K 0.900157440 GB, and F32 0.664721408 GB. The metadata shard records deepseek2.block_count 61, context_length 262144, embedding_length 7168, feed_forward_length 18432, attention.head_count 64, attention.head_count_kv 1, q_lora_rank 1536, kv_lora_rank 512, key_length 576, value_length 512, expert_count 384, expert_used_count 8, expert_shared_count 1, leading_dense_block_count 1, and rope.dimension_count 64." }, { "label": "Audited Moonshot Kimi K2.7 Code profile", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code", "source_type": "manual_review", "supports": [ "base_model_proof", "ordinary_decode_scope", "multimodal" ], "notes": "The existing audited source profile verified the Kimi K2.7 model-card architecture, multimodal wrapper, and ordinary text-decode scope. This GGUF profile differs only in selected artifact weight bytes and the compact GGUF MLA serving-state policy." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, HF CLI model info/card output, repo config.json, the existing audited Moonshot Kimi K2.7 Code profile, linked-object HEAD checks, and direct GGUF header/tensor-index range reads of all eight selected UD-IQ1_M split shards." }, "notes": "Use this profile for the API-selected UD-IQ1_M split GGUF artifact. Do not infer Q2/Q3/Q4/Q8 split-family footprints or multimodal projector residency unless the workload explicitly selects and audits those files." }, { "id": "unsloth--qwen-agentworld-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen-AgentWorld-35B-A3B-GGUF", "title": "Unsloth Qwen AgentWorld 35B A3B GGUF MXFP4_MOE", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth MXFP4_MOE GGUF artifact of Qwen AgentWorld 35B A3B.", "model_family": "qwen3.5-moe-agentworld", "base_model_proof": { "base_model": "Qwen/Qwen-AgentWorld-35B-A3B", "relation": "quantized", "source": "Hugging Face model card/API metadata, Qwen AgentWorld base config, selected linked-object size metadata, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen-AgentWorld-35B-A3B. The selected MXFP4_MOE GGUF header records the same served Qwen3.5 MoE text geometry as the public Qwen AgentWorld config: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 routed experts per token, and a separate always-on shared expert." }, "architecture": { "canonical_architecture_id": "qwen-agentworld-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 21.67049344, "main_resident_weight_gb": 21.119158784, "auxiliary_resident_weight_gb": 0.551334656, "fixed_weight_gb": 2.137836032, "routed_expert_weight_gb": 0.074145792, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected MXFP4_MOE GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "shared_expert_notes": "The model card states 8 routed plus 1 shared expert, and the GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected main GGUF mixes MXFP4, Q5_K, Q6_K, Q8_0, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen AgentWorld text config and GGUF metadata record 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen AgentWorld text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main MXFP4_MOE GGUF artifact. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; multimodal or speculative side paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6252196083637568, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative execution are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen AgentWorld 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen-AgentWorld-35B-A3B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 3a305abf5cfd119ee999dfe929c433746edd8d63 records a public Apache-2.0 text-generation GGUF repo with base_model Qwen/Qwen-AgentWorld-35B-A3B, 349969 downloads, region:us, imatrix metadata, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 21670493440. The API totalFileSize matches Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen AgentWorld 35B A3B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF/raw/3a305abf5cfd119ee999dfe929c433746edd8d63/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "layer_pattern", "routed_experts", "shared_experts_per_token" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model Qwen/Qwen-AgentWorld-35B-A3B, Qwen AgentWorld 35B/3B architecture, 10 x (3 x Gated DeltaNet -> MoE -> 1 x Gated Attention -> MoE), 256 experts, and 8 routed plus 1 shared expert." }, { "label": "Qwen AgentWorld 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B/raw/60d2b0434a53d2e62a7c00a489586815d94ebffb/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The current config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, and 262144 max position embeddings." }, { "label": "Unsloth Qwen AgentWorld 35B A3B GGUF linked-object size checks", "url": "https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF/tree/3a305abf5cfd119ee999dfe929c433746edd8d63", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Expanded HF tree metadata found MXFP4_MOE 21.670493440 GB, Q8_0 36.903140608 GB, UD-IQ2_M 11.563425024 GB, UD-IQ2_XXS 11.496840448 GB, UD-IQ3_S 14.984322304 GB, UD-IQ3_XXS 13.742808320 GB, UD-IQ4_NL 18.112191744 GB, UD-IQ4_XS 17.785036032 GB, UD-Q2_K_XL 12.251290880 GB, UD-Q3_K_M 16.680402176 GB, UD-Q3_K_XL 16.803568896 GB, UD-Q4_K_M 22.134529280 GB, UD-Q4_K_S 20.893015296 GB, UD-Q4_K_XL 22.324804864 GB, UD-Q5_K_M 26.456194304 GB, UD-Q5_K_S 24.942050560 GB, UD-Q5_K_XL 26.527497472 GB, UD-Q6_K 29.308321024 GB, UD-Q6_K_XL 31.843777792 GB, UD-Q8_K_XL 38.201490688 GB, split BF16 total 69.376638464 GB, and imatrix_unsloth.gguf_file 0.192223904 GB. The selected MXFP4_MOE artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Qwen AgentWorld 35B A3B MXFP4_MOE GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF/resolve/3a305abf5cfd119ee999dfe929c433746edd8d63/Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 57 metadata entries and 733 tensors. The linked file is 21.670493440 GB. Tensor spans sum to 21.659503104 GB; metadata/tokenizer/header/file overhead accounts for 0.010990336 GB. Tensor spans split into MXFP4 11.123294208 GB, Q5_K 7.197425664 GB, Q8_0 2.573598720 GB, Q6_K 0.660602880 GB, and F32 0.104581632 GB. token_embd.weight is 0.540344320 GB and resident-only; output.weight plus output_norm.weight is 0.540352512 GB and swept. Routed expert tensors sum to 18.981322752 GB, or 0.074145792 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 2.137836032 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, Qwen AgentWorld base config, expanded linked-file size metadata, a direct GGUF header/tensor-index range read of the API-selected MXFP4_MOE artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected Unsloth Qwen AgentWorld 35B A3B MXFP4_MOE main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen2-5-vl-7b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen2.5-VL-7B-Instruct-GGUF", "title": "Unsloth Qwen2.5 VL 7B Instruct GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen2.5-VL 7B Instruct.", "model_family": "qwen2.5-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen2.5-VL-7B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, shipped Unsloth config comparison, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen2.5-VL-7B-Instruct. The shipped Unsloth config records the same checked architecture fields as the pinned Qwen base config, and the selected BF16 GGUF header records the same qwen2vl text geometry. The target adds Unsloth bookkeeping and GGUF packaging." }, "architecture": { "canonical_architecture_id": "qwen2-5-vl-7b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.615616512, "swept_params_b": 7.070619136, "auxiliary_resident_params_b": 0.544997376, "resident_weight_gb": 15.237851776, "swept_weight_gb": 14.141904896, "auxiliary_resident_weight_gb": 1.09594688, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen2.5-VL-7B-Instruct-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected main GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 15.231899648 GB, while the linked file size is 15.237851776 GB. The main GGUF contains token_embd.weight, blk.* tensors, output.weight, and output_norm.weight. It has no mmproj, vision, visual, audio, MTP, or rope tensor in the selected main file. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Qwen2.5-VL config and selected GGUF metadata record use_sliding_window false, 28 text layers, 4 KV heads, and 128 key/value head dimensions." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000869102585401, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen2.5 VL 7B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen2.5-VL-7B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 68bb8bc4b7df5289c143aaec0ab477a7d4051aab records base_model Qwen/Qwen2.5-VL-7B-Instruct, Apache-2.0 license, image-text-to-text pipeline, region:us, 390164 downloads, GGUF architecture qwen2vl, 128000 context length, gguf.total 7615616512, and gguf.totalFileSize 15237851776." }, { "label": "Unsloth Qwen2.5 VL 7B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen2.5-VL-7B-Instruct, and Qwen2.5-VL multimodal model context." }, { "label": "Unsloth Qwen2.5 VL 7B GGUF config", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/raw/68bb8bc4b7df5289c143aaec0ab477a7d4051aab/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The shipped config records Qwen2_5_VLForConditionalGeneration, tie_word_embeddings false, bfloat16, 28 text layers, hidden size 3584, intermediate size 18944, 28 attention heads, 4 KV heads, sliding_window 32768 but use_sliding_window false, 128000 max position embeddings, vocab size 152064, and a resident vision config. It includes unsloth_fixed true." }, { "label": "Qwen2.5 VL 7B Instruct base config comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/cc594898137f460bfe9f0759e9844b3ce807cfb5/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture" ], "notes": "Manual comparison found no differences in checked architecture, model_type, tie_word_embeddings, dtype, text layer count, hidden size, intermediate size, attention head geometry, sliding-window flags, max_position_embeddings, vocab size, and vision geometry fields between the shipped Unsloth config and the pinned Qwen base config. Differences are limited to Transformers version and Unsloth bookkeeping." }, { "label": "Unsloth Qwen2.5 VL 7B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/tree/68bb8bc4b7df5289c143aaec0ab477a7d4051aab", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen2.5-VL-7B-Instruct-BF16.gguf has linked size 15.237851776 GB, exactly matching API gguf.totalFileSize. Smaller quantized files range from 2.073762688 GB to 10.192153472 GB. Separate mmproj-BF16/F16/F32.gguf sidecars are 1.354163040 GB, 1.354163040 GB, and 2.703221600 GB respectively and are not the selected main artifact." }, { "label": "Unsloth Qwen2.5 VL 7B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/68bb8bc4b7df5289c143aaec0ab477a7d4051aab/Qwen2.5-VL-7B-Instruct-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 96MB range-read of the GGUF v3 header found 28 metadata entries and 339 tensors. The linked file is 15.237851776 GB. Tensor spans sum to 15.231899648 GB: token_embd.weight 1.089994752 GB, blk.* tensors 13.051895808 GB, output.weight 1.089994752 GB, and output_norm.weight 0.000014336 GB. Metadata/tokenizer/header/file overhead accounts for 0.005952128 GB. Stored tensor bytes split into BF16 15.230566400 GB and F32 0.001333248 GB. The header records qwen2vl.block_count 28, context_length 128000, embedding_length 3584, feed_forward_length 18944, attention.head_count 28, attention.head_count_kv 4, rope.dimension_sections [16, 24, 24, 0], and no mmproj, vision, visual, audio, MTP, or rope tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, shipped Unsloth config, pinned Qwen base config comparison, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--qwen3-0-6b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-0.6B-GGUF", "title": "Unsloth Qwen3 0.6B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of Qwen3 0.6B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-0.6B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served config, selected GGUF header metadata, and audited BF16 Qwen3 0.6B profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-0.6B. The served config and selected GGUF header both record Qwen3 geometry matching the audited BF16 base profile: 28 layers, 16 attention heads, 8 KV heads, 128 key/value head dimension, 1024 hidden size, 40960 context, and tied embeddings. The selected GGUF stores token_embd.weight but no separate output.weight, so token_embd.weight is swept as the tied output projection for ordinary decode." }, "architecture": { "canonical_architecture_id": "qwen3-0-6b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.59604992, "swept_params_b": 0.59604992, "auxiliary_resident_params_b": 0, "resident_weight_gb": 1.198182848, "swept_weight_gb": 1.192230912, "auxiliary_resident_weight_gb": 0.005951936, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-0.6B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all stored tensors in the selected GGUF: blk.* tensors, output_norm.weight, and token_embd.weight as the tied output projection", "auxiliary_scope": "GGUF metadata, tokenizer, header, alignment, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected BF16 linked file is 1.198182848 GB. GGUF tensor spans total 1.192230912 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005951936 GB. Because output.weight is not stored separately and the config records tie_word_embeddings true, token_embd.weight is charged as output-projection traffic rather than resident-only input lookup traffic." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 28 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 GGUF artifact. The card lists many lower-bit GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the BF16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Qwen3 0.6B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-0.6B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 50968a4468ef4233ed78cd7c3de230dd1d61a56b, the API records a public non-gated Apache-2.0 GGUF repo with base_model Qwen/Qwen3-0.6B, base_model:quantized, qwen3, unsloth, endpoints_compatible, region:us, 100231 downloads, GGUF architecture qwen3, 40960 context length, gguf.total 596049920, and gguf.totalFileSize 1198182848." }, { "label": "Unsloth Qwen3 0.6B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model Qwen/Qwen3-0.6B, Apache-2.0 license metadata inherited from Qwen, Unsloth GGUF/runtime guidance, and a table of many BF16 and lower-bit GGUF variants." }, { "label": "Unsloth Qwen3 0.6B GGUF served config", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/raw/50968a4468ef4233ed78cd7c3de230dd1d61a56b/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "max_context_tokens", "embedding_layout" ], "notes": "The served config records Qwen3ForCausalLM, qwen3, bfloat16, hidden size 1024, intermediate size 3072, 28 layers, 16 attention heads, 8 KV heads, 128 head dimension, 40960 max position embeddings, use_sliding_window false, tied embeddings, vocab size 151936, and unsloth_fixed true." }, { "label": "Qwen3 0.6B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-0.6B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 28 layers, 8 KV heads, 128 head dimension, 40960 max positions, and a 151936-token vocabulary. It stores a separate lm_head in safetensors, while this selected GGUF omits output.weight and therefore uses tied token_embd.weight for output projection." }, { "label": "Unsloth Qwen3 0.6B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/tree/50968a4468ef4233ed78cd7c3de230dd1d61a56b", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found BF16 1.198182848 GB, Q8_0 0.639447744 GB, Q6_K 0.495107776 GB, Q5_K_M 0.444415680 GB, Q4_K_M 0.396705472 GB, Q3_K_M 0.347127488 GB, Q2_K 0.296238784 GB, and additional Unsloth Dynamic variants. The selected BF16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Qwen3 0.6B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/50968a4468ef4233ed78cd7c3de230dd1d61a56b/Qwen3-0.6B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter", "embedding_layout" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 28 metadata entries and 310 tensors. The linked file is 1.198182848 GB. Tensor spans sum to 1.192230912 GB: token_embd.weight 0.311164928 GB, blk.* tensors 0.881061888 GB, and output_norm.weight 0.000004096 GB. No output.weight tensor is stored. Metadata/tokenizer/header/file overhead accounts for 0.005951936 GB. Stored tensor bytes split into BF16 1.191968768 GB and F32 0.000262144 GB. The header records general.architecture qwen3, qwen3.block_count 28, context_length 40960, embedding_length 1024, feed_forward_length 3072, attention.head_count 16, attention.head_count_kv 8, and key/value head length 128." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, model card, served config, audited BF16 base profile, linked-object HEAD checks for GGUF siblings, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the API-selected Unsloth Qwen3 0.6B BF16 GGUF artifact. Do not copy the existing MaziyarPanahi F16 GGUF split blindly: this selected artifact omits a separate output.weight tensor, so token_embd.weight is swept as tied output projection traffic." }, { "id": "unsloth--qwen3-30b-a3b-instruct-2507-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF", "title": "Unsloth Qwen3 30B A3B Instruct 2507 GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth IQ4_NL GGUF artifact of Qwen3 30B A3B Instruct 2507.", "model_family": "qwen3-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-30B-A3B-Instruct-2507", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, audited BF16 base profile, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card, API metadata, and selected GGUF header identify this package as a GGUF derivative of Qwen/Qwen3-30B-A3B-Instruct-2507. The selected IQ4_NL GGUF header records the same audited Qwen3MoeForCausalLM text geometry as the BF16 base profile: 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 routed experts per token, no shared expert, and 262144 context tokens." }, "architecture": { "canonical_architecture_id": "qwen3-30b-a3b-2507", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.310781856, "main_resident_weight_gb": 17.129781248, "auxiliary_resident_weight_gb": 0.181000608, "fixed_weight_gb": 0.822327296, "routed_expert_weight_gb": 0.127401984, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.47 non-routed tensors, routers, and expected-distinct routed expert tensor groups from the selected IQ4_NL GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "Header-derived stored bytes are used instead of rounded 30B/3B model-card parameters. The selected main GGUF mixes IQ4_NL, Q4_K, Q5_K, Q6_K, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 128 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata and audited BF16 config record 48 full-context attention layers with 4 KV heads and 128 key/value dimensions. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_NL GGUF artifact. The repo does not ship separate projector, MTP, nextn, or draft sidecar tensors for this selected main file." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5669694855212173, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and any runtime-specific expert routing locality are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3 30B A3B Instruct 2507 GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit eea7b2be5805a5f151f8847ede8e5f9a9284bf77 records a public Apache-2.0 GGUF repo with base_model Qwen/Qwen3-30B-A3B-Instruct-2507, 468009 downloads, region:us, imatrix metadata, GGUF architecture qwen3moe, 262144 context length, gguf.total 30532122624, and gguf.totalFileSize 17310781856. The API totalFileSize matches Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3 30B A3B Instruct 2507 GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact" ], "notes": "The card records Apache-2.0 licensing and base_model Qwen/Qwen3-30B-A3B-Instruct-2507. It does not override the API-selected artifact with a specific llama.cpp quantization suffix, so the profile follows gguf.totalFileSize." }, { "label": "Audited BF16 Qwen3 30B A3B Instruct 2507 profile", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The existing audited BF16 profile records Qwen3MoeForCausalLM, qwen3_moe, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, and vocab_size 151936." }, { "label": "Unsloth IQ4_NL GGUF linked object and tensor-index range read", "url": "https://huggingface.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF/resolve/eea7b2be5805a5f151f8847ede8e5f9a9284bf77/Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 45 metadata entries and 579 tensors. The linked file is 17.310781856 GB. Tensor spans sum to 17.304811520 GB; metadata/tokenizer/header/file overhead accounts for 0.005970336 GB. Tensor spans split into IQ4_NL 16.788750336 GB, Q6_K 0.255252480 GB, Q4_K 0.175030272 GB, F32 0.051175424 GB, and Q5_K 0.034603008 GB. token_embd.weight is 0.175030272 GB and resident-only; output.weight is 0.255252480 GB and swept. Routed expert tensors sum to 16.307453952 GB, or 0.127401984 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 0.822327296 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, the existing audited BF16 base profile/config, selected linked file size, and a direct GGUF header/tensor-index range read of the API-selected IQ4_NL artifact." }, "notes": "Use this profile for the API-selected Unsloth Qwen3 30B A3B Instruct 2507 IQ4_NL main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations or runtime-specific expert-cache behavior unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-4B-GGUF", "title": "Unsloth Qwen3 4B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of Qwen3 4B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-4B", "relation": "quantized", "source": "Hugging Face model card/API metadata, pinned repo config, selected GGUF header metadata, and audited BF16 Qwen3 4B base profile", "config_compatible": true, "notes": "The card and API identify this repo as a GGUF derivative of Qwen/Qwen3-4B. The pinned repo config and selected BF16 GGUF header both record the Qwen3 architecture with 36 full-context decoder layers, 32 query heads, 8 KV heads, 128 key/value head dimension, 2560 hidden size, 40960 context, tied embeddings, and no separate output.weight tensor." }, "architecture": { "canonical_architecture_id": "qwen3-4b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.022468096, "swept_params_b": 4.022468096, "auxiliary_resident_params_b": 0, "resident_weight_gb": 8.051285536, "swept_weight_gb": 8.045328384, "auxiliary_resident_weight_gb": 0.005957152, "resident_parameter_scope": "selected GGUF linked file size and API/header logical tensor parameters for Qwen3-4B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all selected BF16 GGUF tensor spans because token_embd.weight is the tied output projection and no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors", "notes": "The API-selected BF16 linked file is 8.051285536 GB. GGUF tensor spans total 8.045328384 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005957152 GB. Because no output.weight tensor is stored, token_embd.weight is charged as ordinary tied output-projection traffic." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The pinned config and selected GGUF header record 36 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 GGUF artifact. Other quantization siblings in the repo should get separate profiles if selected." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2.001578469697824, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is BF16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Qwen3 4B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-4B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 22c9fc8a8c7700b76a1789366280a6a5a1ad1120, the API records a public non-gated Apache-2.0 text-generation GGUF repo with base_model Qwen/Qwen3-4B, base_model:quantized, endpoints_compatible, conversational, region:us, 131101 downloads, GGUF architecture qwen3, context length 40960, gguf.total 4022468096, and gguf.totalFileSize 8051285536. The API totalFileSize matches Qwen3-4B-BF16.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3 4B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/raw/22c9fc8a8c7700b76a1789366280a6a5a1ad1120/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "runtime_format" ], "notes": "The pinned card metadata records base_model Qwen/Qwen3-4B, Apache-2.0 licensing, Transformers library metadata, Qwen3/Unsloth tags, and llama.cpp/Ollama/HF usage guidance for the GGUF package." }, { "label": "Unsloth Qwen3 4B GGUF config", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/raw/22c9fc8a8c7700b76a1789366280a6a5a1ad1120/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The pinned config records Qwen3ForCausalLM, torch_dtype bfloat16, tied embeddings, 36 layers, hidden size 2560, intermediate size 9728, 32 attention heads, 8 KV heads, head dimension 128, 40960 max position embeddings, use_sliding_window false, and no sliding-window setting." }, { "label": "Qwen3 4B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-4B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 36 layers, 8 KV heads, 128 head dimension, 40960 max positions, tied embeddings, and no separate lm_head.weight." }, { "label": "Unsloth Qwen3 4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/22c9fc8a8c7700b76a1789366280a6a5a1ad1120", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found Qwen3-4B-BF16.gguf 8.051285536 GB, Qwen3-4B-Q8_0.gguf 4.280405792 GB, Qwen3-4B-UD-Q4_K_XL.gguf 2.546341152 GB, Qwen3-4B-Q4_K_M.gguf 2.497281312 GB, Qwen3-4B-UD-Q2_K_XL.gguf 1.695724832 GB, and Qwen3-4B-UD-IQ1_M.gguf 1.143484192 GB. The selected BF16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Qwen3 4B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/resolve/22c9fc8a8c7700b76a1789366280a6a5a1ad1120/Qwen3-4B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 28 metadata entries and 398 tensors. The linked file is 8.051285536 GB. Tensor spans sum to 8.045328384 GB: token_embd.weight 0.777912320 GB, blk.* tensors 7.267405824 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.005957152 GB. Stored tensor bytes split into BF16 8.044544000 GB and F32 0.000784384 GB. The header records general.architecture qwen3, qwen3.block_count 36, context_length 40960, embedding_length 2560, feed_forward_length 9728, attention.head_count 32, attention.head_count_kv 8, key/value head length 128, and no output.weight tensor." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned repo config, audited BF16 base profile, linked GGUF file sizes, and direct selected BF16 GGUF tensor-index range read." }, "notes": "Use this profile for the API-selected BF16 GGUF artifact. Do not infer other quantization siblings from this selected-artifact profile." }, { "id": "unsloth--qwen3-5-0-8b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-0.8B-GGUF", "title": "Unsloth Qwen3.5 0.8B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3.5 0.8B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-0.8B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-0.8B. The selected GGUF header records the same Qwen3.5 0.8B text geometry as the Qwen config. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-0-8b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 0.752393024, "swept_params_b": 0.752393024, "auxiliary_resident_params_b": 0, "resident_weight_gb": 1.516744736, "swept_weight_gb": 1.50578304, "auxiliary_resident_weight_gb": 0.010961696, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-0.8B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main artifact; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact file but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 1.505783040 GB, while the linked file size is 1.516744736 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight. It has no output.weight, mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. Because the config records tied embeddings and no output.weight tensor is stored, token_embd.weight remains swept as output-projection traffic for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 24 layers with every fourth layer using full attention, giving 6 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.00132509468881, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and sidecar loading are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3.5 0.8B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-0.8B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 6ab461498e2023f6e3c1baea90a8f0fe38ab64d0 records base_model Qwen/Qwen3.5-0.8B, downloads 283700, Apache-2.0 license metadata, region:us, GGUF architecture qwen35, 262144 context length, gguf.total 752393024, and gguf.totalFileSize 1516744736." }, { "label": "Unsloth Qwen3.5 0.8B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "linear_attention_state", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-0.8B, and Unsloth GGUF packaging with main GGUF artifacts and separate mmproj sidecars." }, { "label": "Qwen3.5 0.8B config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B/raw/2fc06364715b967f1860aea9cf38778875588b17/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, BF16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Unsloth Qwen3.5 0.8B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/tree/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3.5-0.8B-BF16.gguf has linked size 1516744736 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files and mmproj-BF16/F16/F32.gguf sidecars have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Qwen3.5 0.8B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 320 tensors. The linked file is 1.516744736 GB. Tensor spans sum to 1.505783040 GB: token_embd.weight 0.508559360 GB, blk.* tensors 0.997219584 GB, and output_norm.weight 0.000004096 GB. Metadata/tokenizer/header/file overhead accounts for 0.010961696 GB. Stored tensor bytes split into BF16 1.503789056 GB and F32 0.001993984 GB. The header records qwen35.block_count 24, context_length 262144, embedding_length 1024, feed_forward_length 3584, attention.head_count 8, attention.head_count_kv 2, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no output.weight, mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. ggml enum metadata maps tensor type code 30 to BF16." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate files are explicitly loaded by the workload." }, { "id": "unsloth--qwen3-5-27b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-27B-GGUF", "title": "Unsloth Qwen3.5 27B GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth IQ4_NL GGUF artifact of Qwen3.5 27B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD check, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-27B. The selected IQ4_NL GGUF header records the same Qwen3.5 text geometry as the Qwen config, with 64 ordinary text layers and no MTP, mmproj, vision, or draft tensors in the main file." }, "architecture": { "canonical_architecture_id": "qwen3-5-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 15.687894944, "swept_weight_gb": 14.961739776, "auxiliary_resident_weight_gb": 0.726155168, "resident_parameter_scope": "selected Qwen3.5-27B-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected IQ4_NL GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; separate mmproj sidecars are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.5-27B-IQ4_NL.gguf because the live HF API gguf.totalFileSize matches that linked object and the model card does not identify a different normal-serving default. The selected linked file is 15.687894944 GB. Header tensor spans total 15.676901376 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010993568 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, and ordinary blk.0-63 tensors, with no MTP, mmproj, or vision tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_NL GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5832798869689881, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.5-27B-IQ4_NL.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.5 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-27B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 3221f178a6b842d04f1fb42f1c413534adcc0a6a records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.5-27B, 131531 downloads, region:us, imatrix metadata, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 15687894944. The API totalFileSize matches Qwen3.5-27B-IQ4_NL.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3.5 27B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/raw/3221f178a6b842d04f1fb42f1c413534adcc0a6a/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The pinned card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-27B, Qwen3.5 text geometry, and links to Unsloth Dynamic 2.0 GGUF quantization documentation. It does not give a normal-serving llama.cpp command selecting a different quantization, so the API-selected IQ4_NL artifact is used." }, { "label": "Qwen3.5 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 64 text layers, hidden size 5120, intermediate size 17408, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 24 attention heads, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and resident vision config." }, { "label": "Unsloth Qwen3.5 27B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/tree/3221f178a6b842d04f1fb42f1c413534adcc0a6a", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwen3.5-27B-IQ4_NL.gguf 15.687894944 GB, IQ4_XS 14.977484704 GB, Q3_K_M 13.505116064 GB, Q3_K_S 12.289423264 GB, Q4_0 15.721973664 GB, Q4_1 17.182934944 GB, Q4_K_M 16.740812704 GB, Q4_K_S 15.769159584 GB, Q5_K_M 19.608995744 GB, Q5_K_S 18.889000864 GB, Q6_K 22.453933984 GB, Q8_0 28.595763104 GB, UD-IQ2_M 10.188072864 GB, UD-IQ2_XXS 8.573593504 GB, UD-IQ3_XXS 11.506493344 GB, UD-Q2_K_XL 11.213752224 GB, UD-Q3_K_XL 14.438533024 GB, UD-Q4_K_XL 17.621125024 GB, UD-Q5_K_XL 20.171253664 GB, UD-Q6_K_XL 25.675642784 GB, UD-Q8_K_XL 35.528652704 GB, BF16 split parts totaling 53.808281472 GB, imatrix_unsloth.gguf_file 0.013642656 GB, and separate mmproj sidecars of 0.927607040 GB to 1.842940160 GB." }, { "label": "Unsloth Qwen3.5 27B IQ4_NL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/3221f178a6b842d04f1fb42f1c413534adcc0a6a/Qwen3.5-27B-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 49 metadata entries and 851 tensors. The linked file is 15.687894944 GB. Tensor spans sum to 15.676901376 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.715161600 GB, and ordinary blk.0-63 tensors 13.918775296 GB. Metadata/tokenizer/header/file overhead accounts for 0.010993568 GB. Tensor spans split into IQ4_NL 12.787384320 GB, Q5_K 1.095761920 GB, Q6_K 1.042944000 GB, Q4_K 0.715161600 GB, Q8_0 0.025067520 GB, and F32 0.010582016 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 6144, ssm.conv_kernel 4, and no MTP, mmproj, or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, pinned Qwen base config, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected IQ4_NL artifact." }, "notes": "Use this profile for the Unsloth Qwen3.5 normal main IQ4_NL GGUF artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or MTP speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-5-2b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-2B-GGUF", "title": "Unsloth Qwen3.5 2B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3.5 2B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-2B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-2B. The selected GGUF header records the same Qwen3.5 2B text geometry as the Qwen config. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.881825088, "swept_params_b": 1.881825088, "auxiliary_resident_params_b": 0, "resident_weight_gb": 3.775709216, "swept_weight_gb": 3.76474752, "auxiliary_resident_weight_gb": 0.010961696, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-2B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges token_embd.weight as tied output projection plus blk.* tensors and output_norm.weight from the selected main GGUF artifact", "auxiliary_scope": "GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 3.764747520 GB, while the linked file size is 3.775709216 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight, but no output.weight. Because the Qwen3.5 2B config records tied embeddings, token_embd.weight is treated as the tied output projection and remains swept for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 24 layers with every fourth layer using full attention, giving 6 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.019759104, "read_gb_per_output_token": 0.019759104, "state_formula": "18 linear_attention layers * ((16 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 16 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000583127521786, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3.5 2B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-2B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit f6d5376be1edb4d416d56da11e5397a961aca8ae records base_model Qwen/Qwen3.5-2B, Apache-2.0 licensing, region:us, downloads 140054, GGUF architecture qwen35, 262144 context length, gguf.total 1881825088, and gguf.totalFileSize 3775709216." }, { "label": "Unsloth Qwen3.5 2B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "linear_attention_state", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-2B, and the Qwen3.5 architecture summary: 2B language model, hidden size 2048, 24 layers, 6 groups of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 8 query heads, 2 KV heads, 256 attention head dimension, 128-dimensional DeltaNet heads, tied LM output, MTP training, and 262144 native context." }, { "label": "Qwen3.5 2B config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 24 text layers, full_attention_interval 4, 6 full-attention layers, 18 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 16 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, and a resident vision config." }, { "label": "Unsloth Qwen3.5 2B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/tree/f6d5376be1edb4d416d56da11e5397a961aca8ae", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of GGUF siblings found Qwen3.5-2B-BF16.gguf has linked size 3775709216 bytes, exactly matching API gguf.totalFileSize. Nearby alternatives include Q4_K_M at 1280835840 bytes, UD-IQ3_XXS at 931823872 bytes, and mmproj-BF16 at 671372992 bytes, so the API-selected main artifact is the BF16 language GGUF." }, { "label": "Unsloth Qwen3.5 2B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/resolve/f6d5376be1edb4d416d56da11e5397a961aca8ae/Qwen3.5-2B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 320 tensors. The linked file is 3.775709216 GB. Tensor spans sum to 3.764747520 GB: token_embd.weight 1.017118720 GB, blk.* tensors 2.747620608 GB, and output_norm.weight 0.000008192 GB. Metadata/tokenizer/header/file overhead accounts for 0.010961696 GB. Stored tensor bytes split into BF16 3.762552832 GB and F32 0.002194688 GB. The header records qwen35.block_count 24, context_length 262144, embedding_length 2048, feed_forward_length 6144, attention.head_count 8, attention.head_count_kv 2, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 2048, ssm.conv_kernel 4, and no output.weight, mmproj, vision, audio, or MTP tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency, MTP sidecar traffic, or lower-bit sibling footprints unless those separate GGUF files are explicitly loaded and audited by the workload." }, { "id": "unsloth--qwen3-5-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-35B-A3B-GGUF", "title": "Unsloth Qwen3.5 35B A3B GGUF MXFP4_MOE", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth MXFP4_MOE GGUF artifact of Qwen3.5 35B A3B.", "model_family": "qwen3.5-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, served repo config, HF API GGUF metadata, selected linked-object HEAD checks, selected GGUF header metadata, and existing audited Qwen3.5 base config comparison", "config_compatible": true, "notes": "The repo card, API metadata, and served config identify this package as a GGUF derivative of Qwen/Qwen3.5-35B-A3B. The selected MXFP4_MOE GGUF header records the same audited Qwen3.5 text geometry as the Qwen config: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 routed experts per token, and a separate always-on shared expert." }, "architecture": { "canonical_architecture_id": "qwen3-5-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 21.587638912, "main_resident_weight_gb": 21.036304896, "auxiliary_resident_weight_gb": 0.551334016, "fixed_weight_gb": 2.126285312, "routed_expert_weight_gb": 0.073867264, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3.5-35B-A3B-MXFP4_MOE.gguf linked file size including GGUF metadata, tokenizer, header, alignment padding, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected MXFP4_MOE GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; separate mmproj sidecars and imatrix data are not included unless explicitly loaded for another workload", "shared_expert_notes": "The served config records shared_expert_intermediate_size 512, and the selected GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected main GGUF mixes MXFP4, Q5_K, Q8_0, Q6_K, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The served config and selected GGUF metadata record 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main MXFP4_MOE GGUF artifact after any multimodal prefill. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and speculative paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6228291563101036, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, imatrix generation, and speculative execution are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.5-35B-A3B-MXFP4_MOE.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.5 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-35B-A3B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "license", "pipeline", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit bc014a17be43adabd7066b7a86075ff935c6a4e2 records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.5-35B-A3B, 150052 downloads, region:us, endpoints_compatible, imatrix metadata, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 21587638912. The API totalFileSize matches Qwen3.5-35B-A3B-MXFP4_MOE.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3.5 35B A3B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/raw/bc014a17be43adabd7066b7a86075ff935c6a4e2/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "serving" ], "notes": "The pinned card frontmatter records Apache-2.0 licensing, image-text-to-text packaging, and base_model Qwen/Qwen3.5-35B-A3B. The body documents refreshed GGUFs using Unsloth Dynamic 2.0 and new imatrix data." }, { "label": "Unsloth Qwen3.5 35B A3B GGUF config", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/raw/bc014a17be43adabd7066b7a86075ff935c6a4e2/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The served config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, bfloat16 text dtype, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a 27-layer vision config, and one MTP layer in the source config." }, { "label": "Qwen3.5 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/59d61f3ce65a6d9863b86d2e96597125219dc754/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_adapter", "linear_attention_state" ], "notes": "The existing official Qwen3.5 35B A3B audit compared this config against the base model and found matching top-level wrapper, text, vision, MoE, attention, and DeltaNet state geometry. The Unsloth config and selected GGUF header preserve the same memory-relevant text decode layout." }, { "label": "Unsloth Qwen3.5 35B A3B MXFP4_MOE linked object and GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-MXFP4_MOE.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 52 metadata entries and 733 tensors. The linked file is 21.587638912 GB. Tensor spans sum to 21.576649216 GB; metadata/tokenizer/header/file overhead accounts for 0.010989696 GB. Tensor spans split into MXFP4 11.123294208 GB, Q5_K 7.566524416 GB, Q8_0 2.577776640 GB, Q6_K 0.220200960 GB, and F32 0.088852992 GB. token_embd.weight is 0.540344320 GB and resident-only; output.weight is 0.540344320 GB and swept. Routed expert tensors sum to 18.910019584 GB, or 0.073867264 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 2.126285312 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors. HEAD checks found MXFP4_MOE 21.587638912 GB, Q4_K_M 22.016023168 GB, Q4_K_S 20.673845888 GB, BF16 split shards 49.826698560 GB and 19.549939264 GB, mmproj-BF16 0.902822592 GB, mmproj-F16 0.899283648 GB, mmproj-F32 1.786305216 GB, and imatrix sidecar 0.192223904 GB." }, { "label": "GGUF MXFP4 type definitions", "url": "https://github.com/ggml-org/ggml/blob/master/src/ggml-common.h", "source_type": "manual_review", "supports": [ "weight_format", "resident_weight_gb" ], "notes": "Manual review of ggml-common.h found QK_MXFP4 32 and block_mxfp4 as one E8M0 byte plus 16 packed quant bytes per 32 elements." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state plus recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to the value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, served repo config, official Qwen3.5 config evidence, selected linked file sizes, a direct GGUF header/tensor-index range read of the API-selected MXFP4_MOE artifact, ggml MXFP4 type definitions, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected Unsloth Qwen3.5 35B A3B MXFP4_MOE main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, imatrix overhead, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-5-4b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-4B-GGUF", "title": "Unsloth Qwen3.5 4B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3.5 4B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-4B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-4B. The selected GGUF header records the same Qwen3.5 4B text geometry as the Qwen config. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-4b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.205751296, "swept_params_b": 4.205751296, "auxiliary_resident_params_b": 0, "resident_weight_gb": 8.424393632, "swept_weight_gb": 8.413425664, "auxiliary_resident_weight_gb": 0.010967968, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-4B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main artifact; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 8.413425664 GB, while the linked file size is 8.424393632 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight. It has no output.weight, mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. Because the config records tied embeddings and no output.weight tensor is stored, token_embd.weight remains swept as output-projection traffic for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 layers with every fourth layer using full attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0004572481501293, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3.5 4B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-4B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit e87f176479d0855a907a41277aca2f8ee7a09523 records base_model Qwen/Qwen3.5-4B, downloads 1058880, GGUF architecture qwen35, 262144 context length, gguf.total 4205751296, and gguf.totalFileSize 8424393632." }, { "label": "Unsloth Qwen3.5 4B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "linear_attention_state", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-4B, and the Qwen3.5 architecture summary: 4B language model, hidden size 2560, 32 layers, 8 groups of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 16 attention heads, 4 KV heads, 256 attention head dimension, 128-dimensional DeltaNet heads, tied language-model output, MTP training, and 262144 native context." }, { "label": "Qwen3.5 4B config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Unsloth Qwen3.5 4B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/tree/e87f176479d0855a907a41277aca2f8ee7a09523", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3.5-4B-BF16.gguf has linked size 8424393632 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files and mmproj-BF16/F16/F32.gguf sidecars have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Qwen3.5 4B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 426 tensors. The linked file is 8.424393632 GB. Tensor spans sum to 8.413425664 GB: token_embd.weight 1.271398400 GB, blk.* tensors 7.142017024 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.010967968 GB. Stored tensor bytes split into BF16 8.409579520 GB and F32 0.003846144 GB. The header records qwen35.block_count 32, context_length 262144, embedding_length 2560, feed_forward_length 9216, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no output.weight, mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. ggml enum metadata maps tensor type code 30 to BF16." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--qwen3-5-4b-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-4B-MTP-GGUF", "title": "Unsloth Qwen3.5 4B MTP GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3.5 4B with the resident MTP draft block.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-4B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, pinned Qwen base config with MTP settings, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-4B. The selected GGUF header records the same Qwen3.5 4B ordinary text geometry as the Qwen config, plus one extra MTP draft block matching text_config.mtp_num_hidden_layers 1 and mtp_use_dedicated_embeddings false. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-4b-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 4.326350848, "swept_params_b": 4.205751296, "auxiliary_resident_params_b": 0.120599552, "resident_weight_gb": 8.665620544, "swept_weight_gb": 8.413425664, "auxiliary_resident_weight_gb": 0.25219488, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-4B-BF16.gguf in the MTP repo", "swept_parameter_scope": "ordinary text decode charges token_embd.weight, output_norm.weight, and blk.0 through blk.31 tensor spans; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, alignment padding, and blk.32 MTP draft-block tensors are resident in the selected artifact file but not swept for ordinary non-speculative text decode; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 8.654651392 GB, while the linked file size is 8.665620544 GB. The ordinary 32-layer text tensors plus token embedding and output norm match the non-MTP selected BF16 GGUF swept traffic at 8.413425664 GB. The extra MTP draft block blk.32 contributes 0.241225728 GB of resident-only tensor spans for this non-speculative profile. The selected main file has no output.weight, mmproj, vision, audio, or rope_freqs tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 ordinary text layers with every fourth layer using full attention, giving 8 full-context layers. The MTP draft block is not charged as ordinary non-speculative KV traffic in this profile." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the selected main BF16 GGUF artifact after any multimodal prefill. MTP speculative decoding changes the workload and should get a separate profile rather than silently adding draft-block traffic to every ordinary decode token." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000450656007453, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, speculative draft-token execution, and speculative accept/reject behavior are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF with a resident MTP draft block. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3.5 4B MTP GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-4B-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 86835bf9949e4d14d6860f7910b1340ad4f271a9 records base_model Qwen/Qwen3.5-4B, Apache-2.0 image-text-to-text packaging, current downloads 98149, GGUF architecture qwen35, 262144 context length, gguf.total 4326350848, and gguf.totalFileSize 8665620544. The catalog keeps the older qualifying scrape count 115139 until the over-100k working set is regenerated." }, { "label": "Unsloth Qwen3.5 4B MTP GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-MTP-GGUF/raw/86835bf9949e4d14d6860f7910b1340ad4f271a9/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "linear_attention_state", "max_context_tokens", "mtp_scope" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-4B, and MTP speculative decoding guidance using llama.cpp with --spec-type draft-mtp and --spec-draft-n-max 6. It says -np > 1 and --mmproj are not yet supported with MTP. The embedded Qwen overview records a 4B language model, hidden size 2560, 32 ordinary layers, 8 groups of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 16 attention heads, 4 KV heads, 256 attention head dimension, 128-dimensional DeltaNet heads, tied LM output, MTP training, and 262144 native context." }, { "label": "Qwen3.5 4B config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings", "mtp_scope" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings true, bfloat16 text config, 32 ordinary text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, resident vision config, mtp_num_hidden_layers 1, and mtp_use_dedicated_embeddings false." }, { "label": "Unsloth Qwen3.5 4B MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-MTP-GGUF/tree/86835bf9949e4d14d6860f7910b1340ad4f271a9", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3.5-4B-BF16.gguf has linked size 8665620544 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files range from UD-IQ2_M 1.942548800 GB to UD-Q8_K_XL 6.065971520 GB. Separate sidecars are imatrix_unsloth.gguf_file 0.003626496 GB, mmproj-BF16.gguf 0.675569216 GB, mmproj-F16.gguf 0.672423488 GB, and mmproj-F32.gguf 1.334074944 GB; those sidecars are not included in this ordinary text profile." }, { "label": "Unsloth Qwen3.5 4B MTP BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-MTP-GGUF/resolve/86835bf9949e4d14d6860f7910b1340ad4f271a9/Qwen3.5-4B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter", "mtp_scope" ], "notes": "A 64MB range-read of the GGUF v3 header found 44 metadata entries and 441 tensors. The linked file is 8.665620544 GB. Tensor spans sum to 8.654651392 GB: token_embd.weight 1.271398400 GB, blk.* tensors 7.383242752 GB, and output_norm.weight 0.000010240 GB. Metadata/tokenizer/header/file overhead accounts for 0.010969152 GB. Stored tensor bytes split into BF16 8.650752000 GB and F32 0.003899392 GB. Ordinary non-speculative swept tensors are token_embd.weight, output_norm.weight, and blk.0 through blk.31, totaling 8.413425664 GB. The extra blk.32 MTP draft block totals 0.241225728 GB and contains full-attention tensors plus nextn.eh_proj.weight, nextn.enorm.weight, nextn.hnorm.weight, and nextn.shared_head_norm.weight. The header records general.architecture qwen35, qwen35.block_count 33, context_length 262144, embedding_length 2560, feed_forward_length 9216, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, rope.dimension_count 64, and no output.weight, mmproj, vision, audio, or rope_freqs tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, pinned model card, pinned Qwen config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 MTP artifact." }, "notes": "Use this profile for ordinary non-speculative text decode with the Unsloth main BF16 MTP GGUF artifact resident. Do not infer multimodal projector residency, quantized sibling footprints, or speculative MTP throughput unless those separate files and speculative workload semantics are explicitly selected." }, { "id": "unsloth--qwen3-5-9b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-9B-GGUF", "title": "Unsloth Qwen3.5 9B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3.5 9B.", "model_family": "qwen3.5-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-9B. The selected GGUF header records the same Qwen3.5 9B text geometry as the Qwen config. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.953803264, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.01711872, "resident_weight_gb": 17.920697312, "swept_weight_gb": 15.87549184, "auxiliary_resident_weight_gb": 2.045205472, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3.5-9B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors plus output.weight and output_norm.weight from the selected main GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 17.90972928 GB, while the linked file size is 17.920697312 GB. The main GGUF contains token_embd.weight, blk.* tensors, output.weight, and output_norm.weight. It has no mmproj, vision, audio, MTP, or rope_freqs tensor in the selected main file. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 text config and GGUF metadata record 32 layers with every fourth layer using full attention, giving 8 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0002370782490315, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative decoding sidecars are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3.5 9B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-9B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 3885219b6810b007914f3a7950a8d1b469d598a5 records base_model Qwen/Qwen3.5-9B, downloads 1040131, GGUF architecture qwen35, 262144 context length, gguf.total 8953803264, and gguf.totalFileSize 17920697312." }, { "label": "Unsloth Qwen3.5 9B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "linear_attention_state", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.5-9B, and the Qwen3.5 architecture summary: 9B language model, hidden size 4096, 32 layers, 8 groups of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 16 attention heads, 4 KV heads, 256 attention head dimension, 128-dimensional DeltaNet heads, MTP training, and 262144 native context." }, { "label": "Qwen3.5 9B config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, 262144 max position embeddings, a resident vision config, and MTP settings." }, { "label": "Unsloth Qwen3.5 9B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/tree/3885219b6810b007914f3a7950a8d1b469d598a5", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3.5-9B-BF16.gguf has linked size 17920697312 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files and mmproj-BF16/F16/F32.gguf sidecars have different linked sizes and are not the selected main artifact." }, { "label": "Unsloth Qwen3.5 9B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 42 metadata entries and 427 tensors. The linked file is 17.920697312 GB. Tensor spans sum to 17.90972928 GB: token_embd.weight 2.03423744 GB, blk.* tensors 13.841238016 GB, output.weight 2.03423744 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.010968032 GB. Stored tensor bytes split into BF16 17.905483776 GB and F32 0.004245504 GB. The header records qwen35.block_count 32, context_length 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no mmproj/vision/audio/MTP tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or MTP sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--qwen3-5-9b-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.5-9B-MTP-GGUF", "title": "Unsloth Qwen3.5 9B MTP GGUF BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the API-selected BF16 GGUF artifact of the Unsloth Qwen3.5 9B MTP package.", "model_family": "qwen3.5-dense-multimodal-mtp", "base_model_proof": { "base_model": "Qwen/Qwen3.5-9B", "relation": "derived_package", "source": "Hugging Face model card/API metadata, Qwen base config, selected GGUF header metadata, linked-object HEAD checks, and MTP model card guidance", "config_compatible": true, "notes": "The repo card, API metadata, and selected GGUF metadata identify this package as a GGUF derivative of Qwen/Qwen3.5-9B. The package does not ship config.json, so the immutable Qwen config is used for high-level architecture fields. The selected GGUF header records the same Qwen3.5 text geometry plus one resident MTP draft block." }, "architecture": { "canonical_architecture_id": "qwen3-5-9b-mtp", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 9.197093888, "swept_params_b": 7.936684544, "auxiliary_resident_params_b": 1.260409344, "resident_weight_gb": 18.407321728, "swept_weight_gb": 15.87549184, "auxiliary_resident_weight_gb": 2.531829888, "resident_parameter_scope": "selected Qwen3.5-9B-BF16.gguf linked file size and header tensor parameters including ordinary text tensors, input embedding, and the MTP draft block", "swept_parameter_scope": "ordinary non-speculative text decode charges output.weight, output_norm.weight, and blk.0 through blk.31 tensors from the selected BF16 GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.32 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; mmproj sidecar files are not included unless explicitly loaded for another workload", "notes": "The HF API gguf.totalFileSize points at Qwen3.5-9B-BF16.gguf, so this profile targets that selected language-model artifact. The selected linked file is 18.407321728 GB. Header tensor spans total 18.396352512 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010969216 GB. The selected main GGUF contains output.weight, token_embd.weight, output_norm.weight, ordinary blk.0-31 tensors, and blk.32 MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.32 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.5 config and selected GGUF metadata record 32 ordinary text layers with every fourth layer using full attention, giving eight full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.051904512, "read_gb_per_output_token": 0.051904512, "state_formula": "24 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.5 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary non-speculative text decode for the API-selected BF16 GGUF artifact. The MTP draft block is resident for the package but does not reduce token traffic in Bounds Engine v1; multimodal projector use requires a separate workload profile." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0014280545746237, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected profile artifact is the BF16/F32 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.5 9B MTP GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-9B-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "selected_artifact", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 9716a636ee4bddc3fed678220b7a33dd2a4160ae records a public non-gated Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.5-9B, base_model:quantized, endpoints_compatible, region:us, conversational tags, 124435 downloads, GGUF architecture qwen35, 262144 context length, gguf.total 9197093888, and gguf.totalFileSize 18407321728. The API totalFileSize matches Qwen3.5-9B-BF16.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3.5 9B MTP GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/raw/9716a636ee4bddc3fed678220b7a33dd2a4160ae/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "mtp_scope", "auxiliary_scope" ], "notes": "The pinned card metadata records Apache-2.0 licensing, image-text-to-text packaging, and base_model Qwen/Qwen3.5-9B. The card advertises MTP speculative decoding, shows llama.cpp usage with --spec-type draft-mtp and --spec-draft-n-max 6, and states that -np > 1 and --mmproj are not yet supported with MTP in that path. Bounds Engine v1 records the MTP draft block as resident-only for ordinary decode." }, { "label": "Unsloth Qwen3.5 9B MTP GGUF config check", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/raw/9716a636ee4bddc3fed678220b7a33dd2a4160ae/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The pinned package has no config.json at this path, so the profile does not infer architecture from a repo-local config. It uses the immutable Qwen/Qwen3.5-9B config plus selected GGUF metadata." }, { "label": "Qwen3.5 9B config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, bfloat16 text config, 32 text layers, full_attention_interval 4, 8 full-attention layers, 24 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Unsloth Qwen3.5 9B MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/tree/9716a636ee4bddc3fed678220b7a33dd2a4160ae", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found Qwen3.5-9B-BF16.gguf 18.407321728 GB, Q8_0 9.786061152 GB, Q6_K 7.684551008 GB, Q5_K_M 6.729445728 GB, Q4_K_M 5.868826976 GB, UD-Q4_K_XL 6.135034208 GB, UD-Q2_K_XL 4.440797536 GB, UD-IQ2_M 3.969560928 GB, mmproj-BF16 0.921704928 GB, mmproj-F16 0.918165984 GB, and mmproj-F32 1.824061920 GB. The selected BF16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Qwen3.5 9B MTP BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/resolve/9716a636ee4bddc3fed678220b7a33dd2a4160ae/Qwen3.5-9B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 44 metadata entries and 442 tensors. The linked file is 18.407321728 GB. Tensor spans sum to 18.396352512 GB: output.weight 2.034237440 GB, token_embd.weight 2.034237440 GB, ordinary blk.0-31 tensors 13.841238016 GB, blk.32 MTP tensors 0.486623232 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.010969216 GB. Tensor spans split into BF16 18.392023040 GB and F32 0.004329472 GB. The header records qwen35.block_count 33, nextn_predict_layers 1, context_length 262144, embedding_length 4096, feed_forward_length 12288, attention.head_count 16, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.inner_size 4096, and ssm.conv_kernel 4." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, Qwen3.5 base config, missing package-config check, linked-object HEAD checks, direct GGUF header/tensor-index range read of the selected BF16 artifact, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the selected Unsloth Qwen3.5 9B MTP BF16 GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer other quantization bytes, multimodal projector execution, or MTP speculative acceleration unless a workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-6-27b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-27B-GGUF", "title": "Unsloth Qwen3.6 27B GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth IQ4_NL GGUF artifact of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD check, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-27B. The selected IQ4_NL GGUF header records the same Qwen3.6 text geometry as the Qwen config, with 64 ordinary text layers and no MTP, mmproj, vision, or draft tensors in the main file." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 26.895998464, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.2713984, "resident_weight_gb": 16.071772384, "swept_weight_gb": 15.345616896, "auxiliary_resident_weight_gb": 0.726155488, "resident_parameter_scope": "selected Qwen3.6-27B-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected IQ4_NL GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; separate mmproj sidecars are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.6-27B-IQ4_NL.gguf because the live HF API gguf.totalFileSize matches that linked object and the model card does not identify a different normal-serving default. The selected linked file is 16.071772384 GB. Header tensor spans total 16.060778496 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010993888 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, and ordinary blk.0-63 tensors, with no MTP or mmproj tensors." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_NL GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files and MTP package require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5975525469155529, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and write traffic are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.6-27B-IQ4_NL.gguf. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.6 27B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 82d411acf4a06cfb8d9b073a5211bf410bfc29bf records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-27B, 588160 downloads, region:us, imatrix metadata, GGUF architecture qwen35, 262144 context length, gguf.total 26895998464, and gguf.totalFileSize 16071772384. The API totalFileSize matches Qwen3.6-27B-IQ4_NL.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3.6 27B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.6-27B, Qwen3.6 text geometry, and links to Unsloth Dynamic 2.0 GGUF quantization benchmarks. It does not give a normal-serving llama.cpp command selecting a different quantization, so the API-selected IQ4_NL artifact is used." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP training settings." }, { "label": "Unsloth Qwen3.6 27B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/tree/82d411acf4a06cfb8d9b073a5211bf410bfc29bf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HF API sibling metadata records Qwen3.6-27B-IQ4_NL.gguf 16.071772384 GB, IQ4_XS 15.440005344 GB, Q3_K_M 13.586217184 GB, Q3_K_S 12.358727904 GB, Q4_0 15.791278304 GB, Q4_1 17.252239584 GB, Q4_K_M 16.817244384 GB, Q4_K_S 15.856158944 GB, Q5_K_M 19.509790944 GB, Q5_K_S 18.958305504 GB, Q6_K 22.523238624 GB, Q8_0 28.595763424 GB, UD-IQ2_M 10.846136544 GB, UD-IQ2_XXS 9.388779744 GB, UD-IQ3_XXS 11.994777824 GB, UD-Q2_K_XL 11.849779424 GB, UD-Q3_K_XL 14.473431264 GB, UD-Q4_K_XL 17.612564704 GB, UD-Q5_K_XL 20.038256864 GB, UD-Q6_K_XL 25.636485344 GB, UD-Q8_K_XL 35.325163744 GB, BF16 split parts totaling 53.808881760 GB, imatrix_unsloth.gguf_file 0.013642656 GB, and separate mmproj sidecars of 0.927607360 GB to 1.842940480 GB." }, { "label": "Unsloth Qwen3.6 27B IQ4_NL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/resolve/82d411acf4a06cfb8d9b073a5211bf410bfc29bf/Qwen3.6-27B-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the selected GGUF v3 header found 51 metadata entries and 851 tensors. The linked file is 16.071772384 GB. Tensor spans sum to 16.060778496 GB: output.weight plus output_norm.weight 1.042964480 GB, token_embd.weight 0.715161600 GB, and ordinary blk.0-63 tensors 14.302652416 GB. Metadata/tokenizer/header/file overhead accounts for 0.010993888 GB. Tensor spans split into IQ4_NL 11.371806720 GB, Q5_K 2.825912320 GB, Q6_K 1.042944000 GB, Q4_K 0.715161600 GB, and F32 0.104953856 GB. The header records qwen35.block_count 64, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, and no MTP, mmproj, or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned Qwen base config, linked GGUF file sizes, and a direct GGUF header/tensor-index range read of the selected IQ4_NL artifact." }, "notes": "Use this profile for the Unsloth Qwen3.6 normal main IQ4_NL GGUF artifact in ordinary text-decode bounds. Do not infer multimodal projector residency or MTP speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-6-27b-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-27B-MTP-GGUF", "title": "Unsloth Qwen3.6 27B MTP GGUF UD-Q4_K_XL", "summary": "Audited memory-side text-decode bounds profile for the Unsloth card-example UD-Q4_K_XL GGUF artifact of Qwen3.6 27B MTP.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-27B. The selected UD-Q4_K_XL GGUF header records the same Qwen3.6 text geometry as the Qwen config, with 64 ordinary text layers plus one resident MTP draft block." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.320697856, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 1.696097792, "resident_weight_gb": 17.9090976, "swept_weight_gb": 16.916391936, "auxiliary_resident_weight_gb": 0.992705664, "resident_parameter_scope": "logical GGUF parameters from the selected Qwen3.6-27B-UD-Q4_K_XL.gguf header tensor shapes", "swept_parameter_scope": "ordinary text decode charges output.weight, output_norm.weight, and blk.0 through blk.63 tensors from the selected UD-Q4_K_XL GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.64 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; separate mmproj GGUF sidecars are not included unless explicitly loaded for another workload", "notes": "The profile targets Qwen3.6-27B-UD-Q4_K_XL.gguf because the repo card's llama.cpp MTP example uses :UD-Q4_K_XL. The HF API gguf.totalFileSize currently points at Qwen3.6-27B-IQ4_NL.gguf, so this artifact choice is explicit. The selected linked file is 17.9090976 GB. Header tensor spans total 17.898102784 GB, while GGUF metadata, tokenizer, header, and file overhead account for 0.010994816 GB. The selected main GGUF contains output.weight, output_norm.weight, token_embd.weight, ordinary blk.0-63 tensors, and blk.64.nextn MTP tensors. Bounds Engine v1 does not model speculative MTP speedups, so blk.64 is resident-only for ordinary decode." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 64 ordinary layers with every fourth layer using full attention, giving 16 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main UD-Q4_K_XL GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files and the optional MTP speculative execution path require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6555139145564292, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-q4-k-xl-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected UD-Q4_K_XL artifact uses a mixed GGUF layout. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.6 27B MTP GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-27B, 946357 downloads, region:us, imatrix metadata, GGUF architecture qwen35, 262144 context length, gguf.total 27320697856, and gguf.totalFileSize 16337626240. The API totalFileSize matches Qwen3.6-27B-IQ4_NL.gguf; the selected profile uses the card's llama.cpp MTP example artifact instead." }, { "label": "Unsloth Qwen3.6 27B MTP GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.6-27B, Qwen3.6 text geometry, MTP training, and a llama.cpp MTP server command using -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL with draft-mtp enabled. It also notes that multimodal projector loading is not supported together with MTP in the shown llama.cpp path." }, { "label": "Qwen3.6 27B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3_5ForConditionalGeneration, tie_word_embeddings false, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 48 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings." }, { "label": "Unsloth Qwen3.6 27B MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/tree/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of GGUF siblings found Qwen3.6-27B-IQ4_NL.gguf 16.33762624 GB, Qwen3.6-27B-UD-Q4_K_XL.gguf 17.9090976 GB, Qwen3.6-27B-Q4_K_M.gguf 17.10677312 GB, Qwen3.6-27B-Q5_K_M.gguf 19.83405376 GB, Qwen3.6-27B-Q6_K.gguf 22.8844064 GB, Qwen3.6-27B-Q8_0.gguf 29.04708416 GB, BF16 split parts totaling 54.657733728 GB, imatrix_unsloth.gguf_file 0.013642656 GB, and separate mmproj sidecars of 0.92760736 GB to 1.84294048 GB. The selected profile uses UD-Q4_K_XL because it is the artifact in the card's llama.cpp MTP command." }, { "label": "Unsloth Qwen3.6 27B UD-Q4_K_XL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/resolve/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace/Qwen3.6-27B-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 52 metadata entries and 866 tensors. The linked file is 17.9090976 GB. Tensor spans sum to 17.898102784 GB: output.weight 1.042944 GB, output_norm.weight 0.00002048 GB, token_embd.weight 0.7151616 GB, ordinary blk.0-63 tensors 15.873427456 GB, and blk.64 MTP tensors 0.266549248 GB. Metadata/tokenizer/header/file overhead accounts for 0.010994816 GB. Tensor spans split into Q4_K 9.96950016 GB, Q6_K 4.0621056 GB, Q5_K 2.10141184 GB, Q8_0 1.66002688 GB, and F32 0.105058304 GB. The header records qwen35.block_count 65, context_length 262144, embedding_length 5120, feed_forward_length 17408, attention.head_count 24, attention.head_count_kv 4, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, nextn_predict_layers 1, and no mmproj or vision tensor in the selected main file." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from current HF API metadata, model card, pinned Qwen base config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth Qwen3.6 MTP main UD-Q4_K_XL GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer IQ4_NL default serving, multimodal projector residency, or speculative MTP acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-6-27b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-27B-NVFP4", "title": "Unsloth Qwen3.6 27B NVFP4", "summary": "Audited memory-side text-decode bounds profile for the Unsloth NVFP4 package of Qwen3.6 27B.", "model_family": "qwen3.6-dense-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-27B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, NVFP4 recipe model_id, and served config comparison", "config_compatible": true, "notes": "The Unsloth artifact records Qwen/Qwen3.6-27B as its quantized base model. Manual comparison found matching audited top-level, text, and vision geometry fields: Qwen3_5ForConditionalGeneration, 64 text layers, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, one MTP layer, resident vision tower, and 262144 max position embeddings. The Unsloth artifact adds compressed-tensors NVFP4 quantization metadata while preserving the base architecture fields used by this profile." }, "architecture": { "canonical_architecture_id": "qwen3-6-27b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 27.781427952, "swept_params_b": 25.624600064, "auxiliary_resident_params_b": 2.156827888, "resident_weight_gb": 26.380988256, "swept_weight_gb": 22.06733248, "auxiliary_resident_weight_gb": 4.313655776, "resident_parameter_scope": "base logical Qwen3.6 parameters with exact stored NVFP4/BF16/F8/F32 safetensors bytes", "swept_parameter_scope": "model.language_model excluding embed_tokens plus lm_head safetensors headers", "auxiliary_scope": "model.visual tensors, top-level mtp tensors, and model.language_model.embed_tokens.weight are resident for the package but not swept for ordinary text decode", "notes": "Bounds use exact stored bytes from the single safetensors file because the package mixes packed U8 NVFP4 weights, F8_E4M3 scales, F32 global scales, and unquantized BF16 tensors. Logical parameter counts follow the base Qwen3.6 model so model identity remains the 27.781427952B logical architecture while weight traffic follows the quantized artifact bytes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 16, "kv_heads": 4, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 64 layers with every fourth layer marked full_attention, giving 16 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.154927104, "read_gb_per_output_token": 0.154927104, "state_formula": "48 linear_attention layers * ((48 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 48 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The Unsloth artifact preserves the base Qwen3.6 linear-attention geometry and records mamba_ssm_dtype float32. The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers; this profile charges one full fixed-state read per generated token. Compute, writes, and NVFP4 dequantization remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. Image/video prefill throughput is outside this text-decode profile." }, "notes": "Qwen3_5ForConditionalGeneration is multimodal and includes MTP tensors. This profile models ordinary text decode after any multimodal prefill, with speculative MTP disabled." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-qwen3.6-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored NVFP4 packed weights, F8 scales, F32 global scales, and unquantized BF16 tensors from safetensors headers. NVFP4 dequantization, activation traffic, and compute overhead are outside this memory-side bound.", "notes": "The config records compressed-tensors format nvfp4-pack-quantized with 4-bit float weights and local dynamic 4-bit input activations, group_size 16, F8_E4M3 scale dtype, and kv_cache_scheme null. The model card recommends vLLM serving with --dtype bfloat16, so KV cache is charged at two bytes per scalar." }, "evidence": [ { "label": "Unsloth Qwen3.6 27B NVFP4 model card and API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "serving", "total_params_b" ], "notes": "At repo SHA 890bdef7a42feba6d83b6e17a03315c694112f2a, the API records a public Apache-2.0 image-text-to-text repo derived from Qwen/Qwen3.6-27B, with safetensors, unsloth, 8-bit, compressed-tensors, deploy:azure, and region:us tags. Current downloads are 1149369. The API safetensors block reports F32: 608, BF16: 7480996592, F8_E4M3: 1268776960, U8: 10150215680, total: 18899989840. The card describes an Unsloth NVFP4 quantized checkpoint calibrated on UltraChat sequences up to 16K, recommends vLLM serving, and notes that MTP is included but optional for speculative decoding." }, { "label": "Unsloth Qwen3.6 27B NVFP4 config", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/raw/890bdef7a42feba6d83b6e17a03315c694112f2a/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "max_context_tokens", "serving", "auxiliary_resident_scope", "tie_word_embeddings" ], "notes": "The config records Qwen3_5ForConditionalGeneration, language_model_only false, tie_word_embeddings false, compressed-tensors NVFP4 quantization, text_config dtype bfloat16, 64 text layers, full_attention_interval 4, 16 full-attention layers, 48 linear-attention layers, 4 KV heads, 256 full-attention head dimension, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 262144 max position embeddings, resident vision config, and MTP settings. kv_cache_scheme is null." }, { "label": "Qwen3.6 27B base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json", "source_type": "config", "supports": [ "base_model_proof" ], "notes": "Manual comparison found no differences in audited top-level, text_config, layer_types, and vision_config geometry fields between the current base BF16 repo and this Unsloth NVFP4 artifact. The Unsloth artifact adds compressed-tensors quantization_config and recipe metadata but preserves the architecture fields used by this profile." }, { "label": "Unsloth Qwen3.6 27B NVFP4 recipe", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/raw/890bdef7a42feba6d83b6e17a03315c694112f2a/nvfp4_recipe.json", "source_type": "config", "supports": [ "base_model_proof", "weight_format", "quantized_module_scope" ], "notes": "The NVFP4 recipe records model_id Qwen/Qwen3.6-27B, calibration on HuggingFaceH4/ultrachat_200k with max sequence length 16384 and 2000000 token budget, total_linear_modules 607, quantized_linear_modules 304, and ignored_linear_modules 303. Ignored modules include the visual tower, MTP, lm_head, and linear-attention input projections." }, { "label": "Unsloth Qwen3.6 27B NVFP4 safetensors header", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/resolve/890bdef7a42feba6d83b6e17a03315c694112f2a/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_params_b", "swept_params_b", "auxiliary_resident_params_b", "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead" ], "notes": "The single safetensors header was range-read at repo SHA 890bdef7a42feba6d83b6e17a03315c694112f2a. Stored tensor bytes total 26.380988256 GB: BF16 14.961993184 GB, U8 10.150215680 GB, F8_E4M3 1.268776960 GB, and F32 0.000002432 GB. The ordinary text swept subset, defined as model.language_model excluding embed_tokens plus lm_head, totals 22.067332480 GB. The resident-only subset, defined as model.language_model.embed_tokens.weight plus model.visual plus top-level mtp, totals 4.313655776 GB. Stored swept suffix bytes are packed U8 weights 10.150215680 GB, F8 scales 1.268776960 GB, F32 global scales 0.000002432 GB, and unquantized BF16 swept tensors 10.648337408 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from current HF API metadata, the model card, pinned NVFP4 config and recipe, current base config comparison, direct range-read safetensors header, and the Transformers qwen3_5 runtime implementation." }, "notes": "This profile supersedes the scraped metadata estimate, which treated the package as ideal 4-bit dense weights and undercounted unquantized BF16 tensors, F8 scales, F32 scales, MTP, visual, embedding, and output-head storage. It is for ordinary text decode bounds after any multimodal prefill." }, { "id": "unsloth--qwen3-6-35b-a3b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-35B-A3B-GGUF", "title": "Unsloth Qwen3.6 35B A3B GGUF MXFP4_MOE", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth MXFP4_MOE GGUF artifact of Qwen3.6 35B A3B.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD check, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-35B-A3B. The selected MXFP4_MOE GGUF header records the same audited Qwen3.6 text geometry as the Qwen config: 40 text layers, every fourth layer using full attention, 256 routed experts, 8 routed experts per token, and a separate always-on shared expert." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 21.706144736, "main_resident_weight_gb": 21.154810368, "auxiliary_resident_weight_gb": 0.551334368, "fixed_weight_gb": 2.137836032, "routed_expert_weight_gb": 0.074285056, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3.6-35B-A3B-MXFP4_MOE.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected MXFP4_MOE GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token; separate mmproj sidecars are not included unless explicitly loaded for another workload", "shared_expert_notes": "The Qwen model card states 8 routed plus 1 shared expert, and the GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The selected main GGUF mixes MXFP4, Q5_K, Q6_K, Q8_0, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and writes remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main MXFP4_MOE GGUF artifact after any multimodal prefill. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors; sidecars and speculative paths require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.626248190817797, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-mxfp4-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative execution are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3.6-35B-A3B-MXFP4_MOE.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.6 35B A3B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit a483e9e6cbd595906af30beda3187c2663a1118c records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-35B-A3B, 868905 downloads, region:us, imatrix metadata, GGUF architecture qwen35moe, 262144 context length, gguf.total 34660610688, and gguf.totalFileSize 21706144736. The API totalFileSize matches Qwen3.6-35B-A3B-MXFP4_MOE.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3.6 35B A3B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "layer_pattern", "routed_experts", "shared_experts_per_token" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.6-35B-A3B, Qwen3.6 35B/3B architecture, 10 x (3 x Gated DeltaNet -> MoE -> 1 x Gated Attention -> MoE), 256 experts, and 8 routed plus 1 shared expert." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The current config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, tie_word_embeddings false, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, and 262144 max position embeddings." }, { "label": "Unsloth Qwen3.6 35B A3B MXFP4_MOE linked object and GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/resolve/a483e9e6cbd595906af30beda3187c2663a1118c/Qwen3.6-35B-A3B-MXFP4_MOE.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 54 metadata entries and 733 tensors. The linked file is 21.706144736 GB. Tensor spans sum to 21.695154688 GB; metadata/tokenizer/header/file overhead accounts for 0.010990048 GB. Tensor spans split into MXFP4 11.123294208 GB, Q5_K 7.012876288 GB, Q8_0 2.57359872 GB, Q6_K 0.88080384 GB, and F32 0.104581632 GB. token_embd.weight is 0.54034432 GB and resident-only; output.weight is 0.54034432 GB and swept. Routed expert tensors sum to 19.016974336 GB, or 0.074285056 GB per expert index. Fixed ordinary text traffic, including shared experts, routers, attention/DeltaNet tensors, norms, output.weight, and output_norm.weight, sums to 2.137836032 GB. The selected main file contains no mmproj, vision, visual, image, patch, merger, MTP, nextn, or draft tensors." }, { "label": "GGUF MXFP4 type definitions", "url": "https://github.com/ggml-org/ggml/blob/master/src/ggml-common.h", "source_type": "manual_review", "supports": [ "weight_format", "resident_weight_gb" ], "notes": "Manual review of ggml-common.h found QK_MXFP4 32 and block_mxfp4 as one E8M0 byte plus 16 packed quant bytes per 32 elements." }, { "label": "GGML MXFP4 enum definition", "url": "https://github.com/ggml-org/ggml/blob/master/include/ggml.h", "source_type": "manual_review", "supports": [ "weight_format" ], "notes": "Manual review of ggml.h found GGML_TYPE_MXFP4 = 39." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, Qwen base config, selected linked file size, a direct GGUF header/tensor-index range read of the API-selected MXFP4_MOE artifact, ggml MXFP4 type definitions, and the Transformers qwen3_5 runtime implementation." }, "notes": "Use this profile for the API-selected Unsloth Qwen3.6 35B A3B MXFP4_MOE main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations, multimodal projector residency, or speculative acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-6-35b-a3b-mtp-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "title": "Unsloth Qwen3.6 35B A3B MTP GGUF UD-Q4_K_XL", "summary": "Audited memory-side text-decode bounds profile for the Unsloth card-example UD-Q4_K_XL GGUF artifact of Qwen3.6 35B A3B MTP.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, linked-object HEAD checks, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3.6-35B-A3B. The selected UD-Q4_K_XL GGUF header records the same audited Qwen3.6 text geometry as the Qwen config, with 40 ordinary text blocks plus one resident MTP draft block." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 22.853663008, "main_resident_weight_gb": 21.773470208, "auxiliary_resident_weight_gb": 1.0801928, "fixed_weight_gb": 2.137836032, "routed_expert_weight_gb": 0.076701696, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, ordinary blk.0 through blk.39 non-routed tensors, routers, shared experts, and expected-distinct routed expert tensor groups from the selected UD-Q4_K_XL GGUF artifact", "auxiliary_scope": "token_embd.weight, blk.40 MTP draft tensors, and GGUF metadata/tokenizer/header/file overhead are resident in the selected artifact but not swept for ordinary non-speculative text decode; separate mmproj GGUF sidecars are not included unless explicitly loaded for another workload", "shared_expert_notes": "The model card states 8 routed plus 1 shared expert, and the GGUF header records expert_shared_feed_forward_length 512. The v1 adapter folds always-on shared expert traffic into fixed_weight_gb rather than routed expert traffic.", "notes": "The profile targets Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf because the repo card's llama.cpp MTP example uses :UD-Q4_K_XL. The HF API gguf.totalFileSize currently points at Qwen3.6-35B-A3B-MXFP4_MOE.gguf, so this artifact choice is explicit. Routed expert tensors are stored in byte-uniform expert groups across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3.6 text config and GGUF metadata record 40 ordinary layers with every fourth layer using full attention, giving 10 full-context layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The recurrent state uses the post-repeat value-head count and is charged as F32 from the recurrent kernel path; the conv state is charged as FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute, writes, and speculative MTP execution remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main UD-Q4_K_XL GGUF artifact after any multimodal prefill. The separate mmproj GGUF files and optional MTP speculative execution path require separate workload profiles if loaded or enabled." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6436699381307431, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-q4-k-xl-mtp-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, multimodal projector execution, and speculative MTP speedups are outside Bounds Engine v1.", "notes": "The selected UD-Q4_K_XL artifact uses a mixed GGUF layout. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3.6 35B A3B MTP GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 5bc3e238d916f48a861bac2f8a1990a0e9b7e98d records a public Apache-2.0 image-text-to-text GGUF repo with base_model Qwen/Qwen3.6-35B-A3B, 725918 downloads, region:us, imatrix metadata, GGUF architecture qwen35moe, 262144 context length, gguf.total 35505251456, and gguf.totalFileSize 22182574368. The API totalFileSize matches Qwen3.6-35B-A3B-MXFP4_MOE.gguf; the selected profile uses the card's llama.cpp MTP example artifact instead." }, { "label": "Unsloth Qwen3.6 35B A3B MTP GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "auxiliary_scope", "mtp_scope", "routed_experts", "shared_experts_per_token" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, base_model Qwen/Qwen3.6-35B-A3B, Qwen3.6 35B/3B architecture, MTP training, 256 experts, and 8 routed plus 1 shared expert. The llama.cpp MTP server command uses -hf unsloth/Qwen3.6-35B-A3B-MTP-GGUF:UD-Q4_K_XL with --spec-type draft-mtp --spec-draft-n-max 2, and notes that --mmproj is not yet supported with MTP in that path." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable base API commit records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision config, and one MTP layer." }, { "label": "Unsloth Qwen3.6 35B A3B MTP GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/tree/5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of GGUF siblings found MXFP4_MOE 22.182574368 GB, UD-Q4_K_XL 22.853663008 GB, UD-Q4_K_M 22.663387424 GB, UD-Q4_K_S 21.388319008 GB, UD-Q5_K_M 27.087812896 GB, UD-Q5_K_XL 27.159116064 GB, UD-Q6_K 30.011242784 GB, Q8_0 37.801097504 GB, mmproj sidecars from 0.899283584 GB to 1.786305152 GB, and imatrix_unsloth.gguf_file 0.192223904 GB. The selected profile uses UD-Q4_K_XL because it is the artifact in the card's llama.cpp MTP command." }, { "label": "Unsloth Qwen3.6 35B A3B UD-Q4_K_XL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/resolve/5bc3e238d916f48a861bac2f8a1990a0e9b7e98d/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "A 128MB range-read of the GGUF v3 header found 55 metadata entries and 753 tensors. The linked file is 22.853663008 GB. Tensor spans sum to 22.842671616 GB: output.weight plus output_norm.weight 0.540352512 GB, token_embd.weight 0.540344320 GB, ordinary non-routed blk.0-39 tensors 1.597483520 GB, routed expert tensors 19.635634176 GB, and blk.40 MTP tensors 0.528857088 GB. Metadata/tokenizer/header/file overhead accounts for 0.010991392 GB. Tensor spans split into Q4_K 12.079595520 GB, Q5_K 7.381975040 GB, Q8_0 2.614820864 GB, Q6_K 0.660602880 GB, F32 0.104624640 GB, and BF16 0.001052672 GB. The GGUF metadata records qwen35moe.block_count 41, nextn_predict_layers 1, context_length 262144, embedding_length 2048, attention.head_count 16, attention.head_count_kv 2, attention key/value length 256, full_attention_interval 4, ssm.state_size 128, ssm.group_count 16, ssm.conv_kernel 4, expert_count 256, expert_used_count 8, and expert_shared_feed_forward_length 512." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caching conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned Qwen base config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected UD-Q4_K_XL artifact." }, "notes": "Use this profile for the Unsloth Qwen3.6 35B A3B MTP main UD-Q4_K_XL GGUF artifact in ordinary non-speculative text-decode bounds. Do not infer the MXFP4_MOE API default, multimodal projector residency, or speculative MTP acceleration unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-6-35b-a3b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3.6-35B-A3B-NVFP4", "title": "Unsloth Qwen3.6 35B A3B NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for the Unsloth compressed-tensors NVFP4 Qwen3.6 35B A3B artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face base_model metadata, served config comparison, quantization recipe review, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo metadata identifies Qwen/Qwen3.6-35B-A3B as the quantized base. Manual comparison with the current base config found matching checked architecture fields: Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, hidden size 2048, 16 attention heads, 2 KV heads, 256 full-attention head dimension, DeltaNet state geometry, 256 experts, 8 routed experts per token, shared_expert_intermediate_size 512, untied embeddings, 262144 max positions, and resident vision config." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 24.664233168, "main_resident_weight_gb": 21.064690416, "auxiliary_resident_weight_gb": 3.599542752, "fixed_weight_gb": 2.945051376, "routed_expert_weight_gb": 0.07077984, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding, vision tower, and MTP tensors", "auxiliary_scope": "model.visual tensors, mtp tensors, and model.language_model.embed_tokens.weight are resident for the multimodal/MTP package but are not swept for each ordinary generated text token.", "shared_expert_notes": "The model card states 8 routed plus 1 shared expert, and the config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of rounded 35B/3B model-card parameters. The Unsloth recipe targets Linear modules but excludes visual modules, MTP tensors, lm_head, linear-attention projections, MoE router gates, and shared expert gates. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The text config has 40 layers with full_attention_interval 4, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. The Unsloth compressed-tensors NVFP4 artifact preserves the base Qwen3.6 text architecture, so quantizing weights does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as BF16 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The config has kv_cache_scheme null and the model-card vLLM command does not request FP8 KV cache." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower and MTP tensors. This profile models ordinary text decode through the language model and output head, with resident-only multimodal and MTP tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-compressed-tensors-nvfp4-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored compressed-tensors NVFP4 packed weights, FP8 scale tensors, BF16 unquantized tensors, F32 scalar scale tensors, BF16 full-attention KV, and BF16/F32 DeltaNet state bytes. Activation traffic, dequantization, router/expert compute, multimodal encoder compute, MTP speculation, and cache writes are outside this memory-side bound.", "notes": "The repo config records compressed-tensors nvfp4-pack-quantized weights and activations with kv_cache_scheme null. The model card serving command does not request FP8 KV cache, so this profile keeps full-attention KV cache as BF16." }, "evidence": [ { "label": "Unsloth Qwen3.6 35B A3B NVFP4 API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "serving" ], "notes": "At commit 612d523c58522734a1f6dc995bdeebe216647fef, the live API records a public non-gated Apache-2.0 image-text-to-text repo with qwen3_5_moe, unsloth, 8-bit, compressed-tensors, base_model Qwen/Qwen3.6-35B-A3B, and region:us tags. Current downloads are 184698. The API safetensors block records F32 61820, BF16 3089450864, F8_E4M3 2053898240, U8 16431185920, and 21574596844 stored tensor elements." }, { "label": "Unsloth Qwen3.6 35B A3B NVFP4 model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/raw/612d523c58522734a1f6dc995bdeebe216647fef/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "active_params_b", "layers", "kv_heads", "head_dim", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned card describes this as an Unsloth NVFP4 quantized checkpoint calibrated on Hugging Face UltraChat with sequences up to 16K and about a 2M-token calibration budget. The vLLM command uses --dtype bfloat16 and does not request FP8 KV cache. The model overview states 35B total and 3B activated parameters, 40 layers, hidden size 2048, 10 repetitions of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, head dimension 256, 256 experts, 8 routed plus 1 shared expert, 512 expert intermediate dimension, MTP training, and 262144 native context." }, { "label": "Unsloth Qwen3.6 35B A3B NVFP4 config", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/raw/612d523c58522734a1f6dc995bdeebe216647fef/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "quantization_ignore_scope" ], "notes": "The served config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, untied embeddings, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, a resident vision_config, compressed-tensors nvfp4-pack-quantized weights and activations, and kv_cache_scheme null." }, { "label": "Unsloth Qwen3.6 35B A3B NVFP4 quantization recipe", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/raw/612d523c58522734a1f6dc995bdeebe216647fef/nvfp4_recipe.json", "source_type": "config", "supports": [ "weight_format", "calibration", "quantization_ignore_scope" ], "notes": "The pinned recipe records model_id Qwen/Qwen3.6-35B-A3B, text_model_type qwen3_5_moe_text, calibration on HuggingFaceH4/ultrachat_200k train_sft, target 2000000 calibration tokens, 16384 max sequence length, 123 selected 16K calibration samples, resolved_ignore_profile v16_moe, and ignored modules matching lm_head, MTP, visual modules, linear_attn projections, router gates, and shared expert gates. The module summary records 461 total Linear modules, 190 quantized, and 271 ignored." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison against the current base config at commit 995ad96eacd98c81ed38be0c5b274b04031597b0 found matching checked top-level, text_config, and vision_config geometry fields. The Unsloth artifact adds compressed-tensors quantization metadata while preserving the base architecture." }, { "label": "Unsloth Qwen3.6 35B A3B NVFP4 safetensors header", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/resolve/612d523c58522734a1f6dc995bdeebe216647fef/model.safetensors", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Direct range-read of the single safetensors header found 124415 tensors totaling 24.664233168 GB: F32 0.000247280 GB, BF16 6.178901728 GB, F8_E4M3 2.053898240 GB, and U8 16.431185920 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.064690416 GB. Auxiliary resident tensors, defined as model.visual plus mtp plus model.language_model.embed_tokens.weight, sum to 3.599542752 GB. Routed expert tensors sum to 18.119639040 GB and divide exactly into 256 uniform expert indexes of 0.070779840 GB. Fixed ordinary text traffic, including lm_head, DeltaNet weights, full-attention weights, shared experts, gates, and norms, sums to 2.945051376 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caches conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to the value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-07", "notes": "Manual one-model audit from live Unsloth API/card/config/recipe evidence, current Qwen base config comparison, direct safetensors header grouping, and the audited Qwen3.5 runtime state adapter." }, "notes": "This profile supersedes the generated half-byte estimate. It uses exact stored compressed-tensors tensor bytes, charges BF16 full-attention KV plus fixed DeltaNet state, and keeps vision/MTP/input embedding tensors resident-only for ordinary text decode." }, { "id": "unsloth--qwen3-8b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-8B-GGUF", "title": "Unsloth Qwen3 8B GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the API-selected BF16 GGUF artifact of Qwen3 8B.", "model_family": "qwen3-dense-gguf", "base_model_proof": { "base_model": "Qwen/Qwen3-8B", "relation": "quantized", "source": "Hugging Face model card/API metadata, served config, selected GGUF header metadata, and audited BF16 Qwen3 8B base profile", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-8B. The served config and selected GGUF header record the same Qwen3 architecture as the audited BF16 base geometry: 36 layers, 32 attention heads, 8 KV heads, 128 key/value head dimension, 4096 hidden size, 40960 context, and separate input/output embeddings." }, "architecture": { "canonical_architecture_id": "qwen3-8b", "max_context_tokens": 40960, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.19073536, "swept_params_b": 7.568405504, "auxiliary_resident_params_b": 0.622329856, "resident_weight_gb": 16.388044384, "swept_weight_gb": 15.137427456, "auxiliary_resident_weight_gb": 1.250616928, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-8B-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected BF16 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected BF16 linked file is 16.388044384 GB. GGUF tensor spans total 16.382087168 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.005957216 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode. GGUF stores output_norm and layer norms as F32, so tensor bytes are slightly larger than the BF16 safetensors base." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 36 Qwen3 decoder layers with 8 KV heads and 128-dimensional key/value heads. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected BF16 GGUF artifact. The card lists multiple quantized GGUF variants; those should get separate profiles if selected." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2.000802573115975, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for the selected BF16 GGUF artifact. GGUF loader overhead, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The API-selected artifact is the BF16 GGUF. Default llama.cpp-style GGUF KV is modeled as FP16 unless a serving profile explicitly chooses a quantized KV cache." }, "evidence": [ { "label": "Unsloth Qwen3 8B GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-8B-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit a6adef130ffb23ddaf1a62fec9dced968c9bc482, the API records a public non-gated Apache-2.0 GGUF repo with base_model Qwen/Qwen3-8B, base_model:quantized, qwen3, unsloth, endpoints_compatible, region:us, and 110310 downloads. The API GGUF metadata records architecture qwen3, 40960 context length, gguf.total 8190735360, and gguf.totalFileSize 16388044384." }, { "label": "Unsloth Qwen3 8B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/raw/a6adef130ffb23ddaf1a62fec9dced968c9bc482/README.md", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format" ], "notes": "The card records base_model Qwen/Qwen3-8B, Apache-2.0 licensing, Unsloth Qwen3 GGUF packaging, and standard llama.cpp/Ollama-compatible usage guidance. It advertises multiple lower-bit quantizations but does not override the API-selected BF16 artifact with a smaller normal-serving default." }, { "label": "Unsloth Qwen3 8B GGUF config", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/raw/a6adef130ffb23ddaf1a62fec9dced968c9bc482/config.json", "source_type": "config", "supports": [ "base_model_proof", "max_context_tokens", "kv_adapter" ], "notes": "The served config records Qwen3ForCausalLM, qwen3, BF16, hidden size 4096, intermediate size 12288, 36 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 40960, sliding_window null, use_sliding_window false, tie_word_embeddings false, vocab_size 151936, rope_theta 1000000, and unsloth_fixed true." }, { "label": "Qwen3 8B audited base profile", "url": "https://huggingface.co/Qwen/Qwen3-8B", "source_type": "manual_review", "supports": [ "base_model_proof", "embedding_layout", "kv_adapter" ], "notes": "The audited BF16 base profile records the same Qwen3ForCausalLM geometry: 36 layers, 8 KV heads, 128 head dimension, 40960 max positions, and separate stored model.embed_tokens.weight plus lm_head.weight." }, { "label": "Unsloth Qwen3 8B GGUF linked-object checks", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/tree/a6adef130ffb23ddaf1a62fec9dced968c9bc482", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "Range checks found Qwen3-8B-BF16.gguf 16.388044384 GB, Q8_0 8.709519168 GB, Q6_K 6.725900096 GB, Q5_K_M 5.851113280 GB, Q4_K_M 5.027784512 GB, and Q2_K 3.281733440 GB. The selected BF16 artifact exactly matches API gguf.totalFileSize." }, { "label": "Unsloth Qwen3 8B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/resolve/a6adef130ffb23ddaf1a62fec9dced968c9bc482/Qwen3-8B-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 32MB range-read of the selected GGUF v3 header found 28 metadata entries and 399 tensors. The linked file is 16.388044384 GB. Tensor spans sum to 16.382087168 GB: output.weight 1.244659712 GB, token_embd.weight 1.244659712 GB, blk.* tensors 13.892751360 GB, and output_norm.weight 0.000016384 GB. Metadata/tokenizer/header/file overhead accounts for 0.005957216 GB. Stored tensor bytes split into BF16 16.380854272 GB and F32 0.001232896 GB. The header records general.architecture qwen3, qwen3.block_count 36, context_length 40960, embedding_length 4096, feed_forward_length 12288, attention.head_count 32, attention.head_count_kv 8, and key/value head length 128." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, pinned model card, served config, audited BF16 base profile, linked GGUF file sizes, and direct selected BF16 GGUF tensor-index range read." }, "notes": "Use this profile for the Unsloth API-selected Qwen3 8B BF16 GGUF artifact. Do not infer lower-bit sibling footprints unless the workload profile explicitly selects and audits one of those files." }, { "id": "unsloth--qwen3-coder-30b-a3b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "title": "Unsloth Qwen3-Coder 30B A3B Instruct GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth IQ4_NL GGUF artifact of Qwen3-Coder 30B A3B Instruct.", "model_family": "qwen3-coder-moe", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, audited BF16 base profile, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card, API metadata, and selected GGUF header identify this package as a GGUF derivative of Qwen/Qwen3-Coder-30B-A3B-Instruct. The selected IQ4_NL GGUF header records the same audited Qwen3MoeForCausalLM text geometry as the BF16 base profile: 48 layers, 32 attention heads, 4 KV heads, 128 head dimension, 128 routed experts, 8 routed experts per token, no shared expert, and 262144 context tokens." }, "architecture": { "canonical_architecture_id": "qwen3-coder-30b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 17.310784672, "main_resident_weight_gb": 17.129781248, "auxiliary_resident_weight_gb": 0.181003424, "fixed_weight_gb": 0.822327296, "routed_expert_weight_gb": 0.127401984, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.47 non-routed tensors, routers, and expected-distinct routed expert tensor groups from the selected IQ4_NL GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "notes": "Header-derived stored bytes are used instead of rounded 30B/3B model-card parameters. The selected main GGUF mixes IQ4_NL, Q4_K, Q5_K, Q6_K, and F32 tensors. Routed expert tensors are stored in byte-uniform expert groups across all 128 expert indexes." }, "kv_adapter": { "kind": "full_context", "layers": 48, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata and audited BF16 config record 48 full-context attention layers with 4 KV heads and 128 key/value dimensions. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main IQ4_NL GGUF artifact. The repo does not ship separate projector, MTP, nextn, or draft sidecar tensors for this selected main file." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5669695777519487, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, and any runtime-specific expert routing locality are outside Bounds Engine v1.", "notes": "The selected API artifact is Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf. Explicit resident and routed-byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3-Coder 30B A3B Instruct GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit b17cb02dd882d5b6ab62fc777ad2995f19668350, the API records a public Apache-2.0 GGUF repo with base_model Qwen/Qwen3-Coder-30B-A3B-Instruct, transformers library metadata, endpoints_compatible, region:us, imatrix metadata, 239510 downloads, GGUF architecture qwen3moe, 262144 context length, gguf.total 30532122624, and gguf.totalFileSize 17310784672. The API totalFileSize matches Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3-Coder 30B A3B Instruct GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact", "model_family" ], "notes": "The card records Apache-2.0 licensing and base_model Qwen/Qwen3-Coder-30B-A3B-Instruct. It documents 30.5B total parameters, 3.3B activated parameters, 48 layers, 32 Q heads, 4 KV heads, 128 experts, 8 activated experts, 262144 native context, and non-thinking Qwen3-Coder behavior. It does not override the API-selected artifact with a specific llama.cpp quantization suffix, so this profile follows gguf.totalFileSize." }, { "label": "Audited BF16 Qwen3-Coder 30B A3B Instruct profile", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json", "source_type": "manual_review", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "max_context_tokens" ], "notes": "The existing audited BF16 profile records Qwen3MoeForCausalLM, qwen3_moe, 48 layers, hidden_size 2048, 32 attention heads, 4 KV heads, head_dim 128, 128 experts, 8 experts per token, no shared expert, decoder_sparse_step 1, sliding_window null, use_sliding_window false, tie_word_embeddings false, max_position_embeddings 262144, vocab_size 151936, and rope_theta 10000000." }, { "label": "Unsloth Qwen3-Coder 30B A3B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/tree/b17cb02dd882d5b6ab62fc777ad2995f19668350", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "HEAD checks found IQ4_NL 17.310784672 GB, IQ4_XS 16.378076320 GB, Q2_K 11.258612896 GB, Q3_K_M 14.711850144 GB, Q4_K_M 18.556689568 GB, Q5_K_M 21.725584544 GB, Q6_K 25.092535456 GB, Q8_0 32.483935392 GB, UD-Q4_K_XL 17.665334432 GB, BF16 split parts totaling 61.095306048 GB, and imatrix_unsloth.gguf_file 0.122029344 GB. The API-selected IQ4_NL linked size exactly matches gguf.totalFileSize." }, { "label": "Unsloth Qwen3-Coder 30B A3B IQ4_NL GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/resolve/b17cb02dd882d5b6ab62fc777ad2995f19668350/Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 44 metadata entries and 579 tensors. The linked file is 17.310784672 GB. Tensor spans sum to 17.304811520 GB; metadata/tokenizer/header/file overhead accounts for 0.005973152 GB. Tensor spans split into IQ4_NL 16.788750336 GB, Q6_K 0.255252480 GB, Q4_K 0.175030272 GB, F32 0.051175424 GB, and Q5_K 0.034603008 GB. token_embd.weight is 0.175030272 GB and resident-only; output.weight is 0.255252480 GB and swept. Routed expert tensors sum to 16.307453952 GB, or 0.127401984 GB per expert index. Fixed ordinary text traffic, including routers, attention tensors, norms, output.weight, and output_norm.weight, sums to 0.822327296 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, the existing audited BF16 base profile/config, selected linked file size, HEAD checks for GGUF siblings, and a direct GGUF header/tensor-index range read of the API-selected IQ4_NL artifact." }, "notes": "Use this profile for the API-selected Unsloth Qwen3-Coder 30B A3B Instruct IQ4_NL main GGUF artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations or runtime-specific expert-cache behavior unless the workload profile explicitly selects those paths." }, { "id": "unsloth--qwen3-coder-next-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-Coder-Next-GGUF", "title": "Unsloth Qwen3-Coder-Next GGUF IQ4_NL", "summary": "Audited memory-side text-decode bounds profile for the API-selected Unsloth IQ4_NL GGUF artifact of Qwen3-Coder-Next.", "model_family": "qwen3-next-moe-coder", "base_model_proof": { "base_model": "Qwen/Qwen3-Coder-Next", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, selected linked-object HEAD check, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-Coder-Next. The selected GGUF header records qwen3next architecture, 48 blocks, 512 experts, 10 active experts, one shared expert, full_attention_interval 4, Qwen3Next SSM metadata, and 262144 context, matching the audited Qwen BF16 config on checked memory-relevant fields." }, "architecture": { "canonical_architecture_id": "qwen3-coder-next", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 45.13034448, "main_resident_weight_gb": 44.9493248, "auxiliary_resident_weight_gb": 0.18101968, "fixed_weight_gb": 1.462780928, "routed_expert_weight_gb": 0.084934656, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected Qwen3-Coder-Next-IQ4_NL.gguf linked file size including GGUF metadata, tokenizer, header, alignment padding, and tensor spans", "traffic_scope": "ordinary text decode charges output.weight, output_norm.weight, all non-input-embedding blk.0 through blk.47 non-routed tensors, shared expert tensors, routers, and expected-distinct routed expert tensor groups from the selected IQ4_NL GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected main artifact but not swept for each ordinary text decode token", "shared_expert_notes": "The model card, base config, and selected GGUF metadata record one shared expert. Shared expert and shared expert gate tensors are outside blk.*.ffn_*_exps.weight and are charged in fixed_weight_gb as always-on MoE-layer traffic.", "notes": "The profile targets Qwen3-Coder-Next-IQ4_NL.gguf because HF API gguf.totalFileSize exactly matches that linked object. GGUF tensor spans total 45.124355072 GB, while metadata/header/alignment padding accounts for 0.005989408 GB. Routed expert tensors total 43.486543872 GB and divide exactly by 512 expert indexes. Non-expert tensor spans total 1.637811200 GB; after treating token_embd.weight as resident-only, fixed ordinary decode traffic is 1.462780928 GB." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The Qwen3Next config and selected GGUF metadata record full_attention_interval 4 over 48 layers, giving 12 full-attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.077856768, "read_gb_per_output_token": 0.077856768, "state_formula": "36 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 FP16 bytes))", "notes": "The Qwen3Next implementation caches conv_state plus recurrent_state for linear-attention layers. The gated-delta recurrent kernels cast Q/K/V/B/G to torch.float32 before producing last_recurrent_state, while conv_state is charged as activation-side FP16 for GGUF serving. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3-Next text decode is represented as full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the API-selected main IQ4_NL GGUF artifact. Other GGUF quantizations in the same repo have different resident and traffic bytes and require separate workload selection if used." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.5664347570894557, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-iq4-nl-qwen3next-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, kernels, dequantization, scheduler behavior, tool-calling templates, and state writes are outside Bounds Engine v1.", "notes": "The selected API artifact is an Unsloth IQ4_NL GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected linked file over logical GGUF parameters." }, "evidence": [ { "label": "Unsloth Qwen3-Coder-Next GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-Coder-Next-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit ce09c67b53bc8739eef83fe67b2f5d293c270632 records a public non-gated Apache-2.0 text-generation GGUF repo with base_model Qwen/Qwen3-Coder-Next, 233649 downloads, region:us, GGUF architecture qwen3next, 262144 context length, gguf.total 79674391296, and gguf.totalFileSize 45130344480. The API totalFileSize matches Qwen3-Coder-Next-IQ4_NL.gguf, so this profile targets that artifact." }, { "label": "Unsloth Qwen3-Coder-Next GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "serving", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "layer_pattern", "max_context_tokens" ], "notes": "The pinned card identifies Qwen/Qwen3-Coder-Next as the quantized base, recommends more than 45GB unified memory/RAM/VRAM for 4-bit quants, and records 80B total parameters, 3B activated parameters, 48 layers in a 12 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE)) layout, 16 Q heads and 2 KV heads, 32 linear V heads and 16 linear QK heads, 512 experts, 10 active experts, one shared expert, and 262144 native context." }, { "label": "Qwen3-Coder-Next base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "full_attention_interval", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "tie_word_embeddings" ], "notes": "The base config records Qwen3NextForCausalLM, qwen3_next, bfloat16 runtime dtype, untied embeddings, 48 layers, full_attention_interval 4, hidden size 2048, 512 routed experts, 10 routed experts per token, shared_expert_intermediate_size 512, 16 attention heads, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, and 262144 max position embeddings." }, { "label": "Unsloth Qwen3-Coder-Next IQ4_NL GGUF linked object and range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF/resolve/ce09c67b53bc8739eef83fe67b2f5d293c270632/Qwen3-Coder-Next-IQ4_NL.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "layers", "kv_heads", "head_dim", "linear_attention_state", "weight_format", "kv_adapter" ], "notes": "HF linked-object metadata reports 45.130344480 GB for Qwen3-Coder-Next-IQ4_NL.gguf, matching API gguf.totalFileSize. A direct GGUF v3 range-read found 53 metadata entries and 843 tensors. Tensor spans sum to 45.124355072 GB across 79.674391296B logical elements; metadata/header/alignment padding accounts for 0.005989408 GB. Tensor spans split into IQ4_NL 44.175753216 GB, Q4_0 0.296540160 GB, Q4_1 0.216268800 GB, IQ3_XXS 0.053477376 GB, F32 0.207285248 GB, and Q8_0 0.175030272 GB. token_embd.weight is 0.175030272 GB and resident-only; output.weight is 0.255252480 GB and swept. Routed expert tensors sum to 43.486543872 GB, or 0.084934656 GB per expert index. Fixed ordinary text traffic, including attention tensors, linear-attention tensors, shared experts, routers, norms, output.weight, and output_norm.weight, sums to 1.462780928 GB. The header records qwen3next.block_count 48, context_length 262144, expert_count 512, expert_used_count 10, full_attention_interval 4, attention.head_count 16, attention.head_count_kv 2, key/value length 256, ssm.group_count 16, ssm.inner_size 4096, ssm.state_size 128, and ssm.conv_kernel 4." }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 9f909e0f7027285d83bb6addc88155f1b80244ab found Qwen3NextConfig generating full_attention every fourth layer, DynamicCache dispatching full_attention to K/V cache and linear_attention to LinearAttentionLayer, and Qwen3NextGatedDeltaNet caching conv_states and recurrent_states. The gated-delta kernels cast Q/K/V/B/G to torch.float32 and create last_recurrent_state with shape [batch, repeated value heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, Qwen base config, selected linked file size, a direct GGUF header/tensor-index range read of the selected IQ4_NL artifact, and the Transformers qwen3_next runtime implementation." }, "notes": "Use this profile for the API-selected Unsloth IQ4_NL GGUF main artifact in ordinary text-decode bounds. Do not infer other GGUF quantizations from this profile unless the workload explicitly selects those files." }, { "id": "unsloth--qwen3-vl-2b-instruct-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Qwen3-VL-2B-Instruct-GGUF", "title": "Unsloth Qwen3 VL 2B Instruct GGUF BF16", "summary": "Audited memory-side text-decode bounds profile for the selected Unsloth BF16 GGUF artifact of Qwen3-VL 2B Instruct.", "model_family": "qwen3-vl-dense", "base_model_proof": { "base_model": "Qwen/Qwen3-VL-2B-Instruct", "relation": "quantized", "source": "Hugging Face model card base_model metadata, HF API GGUF metadata, Qwen base config, and GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of Qwen/Qwen3-VL-2B-Instruct. The selected BF16 GGUF header records the same Qwen3-VL text geometry as the Qwen config. The Unsloth GGUF repo does not ship config.json, so the immutable Qwen config is used for high-level architecture fields." }, "architecture": { "canonical_architecture_id": "qwen3-vl-2b", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.720574976, "swept_params_b": 1.720574976, "auxiliary_resident_params_b": 0, "resident_weight_gb": 3.44735072, "swept_weight_gb": 3.44139776, "auxiliary_resident_weight_gb": 0.00595296, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for Qwen3-VL-2B-Instruct-BF16.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main artifact; token_embd.weight is the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; separate mmproj GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the BF16 GGUF file selected by HF API gguf.totalFileSize. Header tensor spans total 3.441397760 GB, while the linked file size is 3.447350720 GB. The main GGUF contains token_embd.weight, blk.* tensors, and output_norm.weight. It has no output.weight, mmproj, vision, visual, audio, MTP, or draft tensor in the selected main file. Because the config records tied embeddings and no output.weight tensor is stored, token_embd.weight remains swept as output-projection traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The Qwen3-VL text config and selected GGUF metadata record full-context attention geometry with 28 layers, 8 KV heads, and 128 key/value head dimensions. The served config does not define a sliding-window or recurrent-state text cache." }, "notes": "This profile models ordinary text decode for the selected main BF16 GGUF artifact after any multimodal prefill. The separate mmproj-BF16/F16/F32 GGUF files require separate workload profiles if loaded." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0001440262722965, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-bf16-f32-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for weight traffic and selected linked artifact size for residency. GGUF loader overhead, kernels, scheduler behavior, multimodal projector execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the Unsloth BF16 GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes the selected artifact's BF16 plus small F32 tensor storage, excluding GGUF header/tokenizer overhead." }, "evidence": [ { "label": "Unsloth Qwen3 VL 2B GGUF HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-VL-2B-Instruct-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 8dcb98e52a1d1d02dce9249e5ab15bae8121c666 records base_model Qwen/Qwen3-VL-2B-Instruct, Apache-2.0 license, image-text-to-text pipeline, region:us, 856615 downloads, GGUF architecture qwen3vl, 262144 context length, gguf.total 1720574976, and gguf.totalFileSize 3447350720." }, { "label": "Unsloth Qwen3 VL 2B GGUF model card", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, image-text-to-text packaging, and base_model Qwen/Qwen3-VL-2B-Instruct." }, { "label": "Qwen3 VL 2B Instruct config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/89644892e4d85e24eaac8bacfd4f463576704203/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The immutable config records Qwen3VLForConditionalGeneration, tie_word_embeddings true, BF16 text dtype, 28 text layers, hidden_size 2048, 16 attention heads, 8 KV heads, 128 head dimension, 262144 max position embeddings, vocab_size 151936, Interleaved-MRoPE sections, and a resident 24-layer visual tower with hidden_size 1024." }, { "label": "Unsloth Qwen3 VL 2B GGUF linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF/tree/8dcb98e52a1d1d02dce9249e5ab15bae8121c666", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks of all GGUF siblings found Qwen3-VL-2B-Instruct-BF16.gguf has linked size 3447350720 bytes, exactly matching API gguf.totalFileSize. Smaller quantized files range from 537831104 to 2332583616 bytes. Separate mmproj-BF16/F16/F32.gguf sidecars are 822540960, 819395232, and 1627847328 bytes respectively and are not the selected main artifact." }, { "label": "Unsloth Qwen3 VL 2B BF16 GGUF range-read tensor index", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF/resolve/8dcb98e52a1d1d02dce9249e5ab15bae8121c666/Qwen3-VL-2B-Instruct-BF16.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 38 metadata entries and 310 tensors. The linked file is 3.447350720 GB. Tensor spans sum to 3.441397760 GB: token_embd.weight 0.622329856 GB, blk.* tensors 2.819059712 GB, and output_norm.weight 0.000008192 GB. Metadata/tokenizer/header/file overhead accounts for 0.005952960 GB. Stored tensor bytes split into BF16 3.440902144 GB and F32 0.000495616 GB. The header records qwen3vl.block_count 28, context_length 262144, embedding_length 2048, feed_forward_length 6144, attention.head_count 16, attention.head_count_kv 8, attention key/value length 128, rope.dimension_sections [24, 20, 20, 0], n_deepstack_layers 3, and no output.weight, mmproj, vision, visual, audio, MTP, or draft tensor in the selected main file. ggml enum metadata maps tensor type code 30 to BF16." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, pinned Qwen config, HEAD checks for all GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected BF16 artifact." }, "notes": "Use this profile for the Unsloth main BF16 GGUF text artifact. Do not infer multimodal projector residency or sidecar traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "unsloth--step-3-7-flash-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "unsloth/Step-3.7-Flash-GGUF", "title": "Unsloth Step 3.7 Flash GGUF UD-Q4_K_S", "summary": "Audited memory-side text-decode bounds profile for the local-serving UD-Q4_K_S GGUF split artifact of Step 3.7 Flash.", "model_family": "step3p7-moe-vlm-gguf", "base_model_proof": { "base_model": "stepfun-ai/Step-3.7-Flash", "relation": "quantized", "source": "Hugging Face API base_model metadata, Unsloth model card, pinned StepFun base config, and direct GGUF split header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of stepfun-ai/Step-3.7-Flash. The selected split GGUF metadata records the same ordinary text geometry used by the StepFun base config: 45 text blocks, 262144 context, 12 full-attention layers, 33 sliding-window layers, 512-token sliding window, 8 KV groups, 128 key/value length, 288 routed experts, 8 experts per token, one shared expert, and three leading dense blocks." }, "architecture": { "canonical_architecture_id": "step-3-7-flash-llamacpp-gguf", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 114.163192448, "main_resident_weight_gb": 113.596957184, "auxiliary_resident_weight_gb": 0.566235264, "fixed_weight_gb": 6.579290624, "routed_expert_weight_gb": 0.37158912, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "selected four-part UD-Q4_K_S GGUF linked file sizes including metadata-only first split, tokenizer, tensor spans, and GGUF overhead", "traffic_scope": "ordinary llama.cpp text decode through the selected main text GGUF split package, charging output.weight, output_norm.weight, rope frequencies, all non-routed blk.0 through blk.44 tensors, routers, shared experts, dense early layers, and expected-distinct routed expert tensors", "auxiliary_scope": "token_embd.weight plus GGUF metadata/tokenizer/header/file overhead are resident in the selected split package but not swept as full matrices for each ordinary generated token; the separate mmproj GGUF sidecars are not included unless explicitly loaded for a multimodal workload", "shared_expert_notes": "The base config and GGUF metadata record one shared expert branch per MoE layer. Shared expert tensors are stored as blk.*.ffn_*_shexp.weight and are included in fixed_weight_gb because they are always read when a MoE layer runs.", "notes": "Layers 0-2 are dense FFN layers and layers 3-44 are MoE layers. The selected GGUF stores routed experts as fused per-layer tensors named blk.*.ffn_down_exps.weight, blk.*.ffn_gate_exps.weight, and blk.*.ffn_up_exps.weight. Those routed tensors are byte-uniform across the 288 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 12, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata marks layers 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, and 44 as full-attention layers." }, { "kind": "sliding_window", "layers": 33, "kv_heads": 8, "head_dim": 128, "window_tokens": 512, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata marks the remaining 33 layers as sliding-window layers with step35.attention.sliding_window 512." } ], "notes": "The selected GGUF package does not declare a required quantized KV-cache format, so the llama.cpp-style serving bound charges FP16 K/V cache storage and read traffic. Vision prefill, mmproj sidecar execution, and speculative/MTP execution are outside this ordinary text-decode profile." }, "notes": "This profile targets the main text UD-Q4_K_S split package for local serving. The separate mmproj-BF16/F16/F32 sidecar files are needed for multimodal image workloads but are not ordinary text decode traffic." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.579637669554111, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-ud-q4-k-s-step3p7-moe-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and selected linked split sizes for residency. GGUF loader overhead, tokenizer work, kernels, dequantization, router compute, scheduler behavior, multimodal projector execution, and speculative decoding are outside Bounds Engine v1.", "notes": "The selected UD-Q4_K_S split stores routed expert tensors as Q4_K, token embedding and most side tensors as Q8_0, output.weight as Q6_K, and small norms/routers/biases as F32. Default GGUF KV is modeled as FP16 unless a serving profile explicitly chooses quantized KV." }, "evidence": [ { "label": "Unsloth Step 3.7 Flash GGUF API metadata", "url": "https://huggingface.co/api/models/unsloth/Step-3.7-Flash-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "pipeline", "base_model_proof", "license", "max_context_tokens", "total_params_b" ], "notes": "At commit 8dd62fe8a98f3c8994160d0c521542560b558f37, the live API records a public non-gated Apache-2.0 image-text-to-text GGUF repo with transformers, vision-language, unsloth multimodal MoE, base_model:stepfun-ai/Step-3.7-Flash, endpoints_compatible, region:us, and conversational tags. Current downloads are 44668; the catalog row keeps the original qualifying scrape count 105722. The API GGUF block records architecture step35, context_length 262144, logical tensor elements 196956130432, and totalFileSize 394017424160 for the BF16 aggregate." }, { "label": "Unsloth Step 3.7 Flash GGUF model card", "url": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF/raw/8dd62fe8a98f3c8994160d0c521542560b558f37/README.md", "source_type": "model_card", "supports": [ "selected_artifact", "base_model_proof", "architecture", "serving" ], "notes": "The card describes Step 3.7 Flash as a 198B sparse MoE VLM with a 196B language backbone, 1.8B vision encoder, about 11B activated parameters per token, and a 256k context window. Its llama.cpp section lists Q4_K_S 111.5 GB, IQ4_XS 104.99 GB, an FP16 multimodal projector around 3.97 GB, 128 GB recommended unified memory, and run/benchmark commands using Step3.7_Q4_K_S.gguf. Although the API totalFileSize matches the BF16 split, this local-serving guidance selects UD-Q4_K_S for this local-frontier profile." }, { "label": "Step 3.7 Flash base config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter" ], "notes": "The pinned base config records Step3p7ForConditionalGeneration with a Step3p5 text_config: 45 text layers, 262144 max positions, 512-token sliding window, 64 full-attention query heads, 96 sliding-attention query heads through layer_types, 8 attention groups, head_dim 128, moe_num_experts 288, moe_top_k 8, MoE on layers 3-44 via three leading dense blocks in the GGUF metadata, share_expert_dim 1280, and num_nextn_predict_layers 3." }, { "label": "Unsloth Step 3.7 Flash UD-Q4_K_S split GGUF header audit", "url": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF/tree/8dd62fe8a98f3c8994160d0c521542560b558f37/UD-Q4_K_S", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "weight_format", "kv_adapter" ], "notes": "Direct HEAD checks and GGUF v3 range reads of the selected four UD-Q4_K_S split files found sizes 0.005232064 GB, 49.972080128 GB, 49.973587520 GB, and 14.212292736 GB, totaling 114.163192448 GB. The first split is metadata-only and records step35 block_count 45, context_length 262144, 12 full-attention layers, 33 sliding-window layers, sliding_window 512, attention.head_count_kv all 8, attention key/value length 128, expert_count 288, expert_used_count 8, expert_feed_forward_length 1280, expert_shared_feed_forward_length 1280, leading_dense_block_count 3, and split.tensors.count 754. Tensor spans across splits 2-4 sum to 114.157912576 GB, with 0.005279872 GB GGUF metadata/header/file overhead. Tensor spans split into Q4_K 107.017666560 GB, Q8_0 6.507388928 GB, Q6_K 0.433090560 GB, and F32 0.199766528 GB. token_embd.weight is 0.560955392 GB and resident-only. Routed expert tensors total 107.017666560 GB, or 0.371589120 GB per expert index. Fixed ordinary text traffic, including output.weight, output_norm.weight, rope_freqs, dense layers, attention tensors, routers, shared experts, norms, and small biases, is 6.579290624 GB." }, { "label": "Selected UD-Q4_K_S artifact", "url": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF/resolve/8dd62fe8a98f3c8994160d0c521542560b558f37/UD-Q4_K_S/Step-3.7-Flash-UD-Q4_K_S-00001-of-00004.gguf", "source_type": "derived_calculation", "supports": [ "selected_artifact", "serving" ], "notes": "The selected package consists of Step-3.7-Flash-UD-Q4_K_S-00001-of-00004.gguf through Step-3.7-Flash-UD-Q4_K_S-00004-of-00004.gguf. The selected split contains no vision, visual, image, mmproj, MTP, or NextN tensors; multimodal projectors are separate GGUF sidecars." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current Hugging Face API metadata, pinned Unsloth model card, pinned StepFun base config, linked-object HEAD checks, and direct GGUF header/tensor-index range reads across all four selected UD-Q4_K_S split files." }, "notes": "Use this profile for local ordinary text decode with the Unsloth UD-Q4_K_S Step 3.7 Flash GGUF package. Do not silently substitute the BF16 API aggregate, the MXFP4_MOE sibling, the UD-Q4_K_XL sibling, or the separate mmproj files; those require their own selected-artifact profiles." }, { "id": "vlsav--vikhr-nemo-12b-instruct-r-21-09-24-q8-0-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF", "title": "VlSav Vikhr-Nemo 12B Instruct GGUF Q8_0", "summary": "Audited memory-side text-decode bounds profile for the single Q8_0 GGUF artifact of Vikhr-Nemo 12B Instruct.", "model_family": "mistral-nemo-dense-gguf", "base_model_proof": { "base_model": "Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24", "relation": "quantized", "source": "Hugging Face API metadata, model card, original Vikhr config/API metadata, selected GGUF header metadata, and Mistral Nemo base config comparison", "config_compatible": true, "notes": "The GGUF repo card and API metadata identify this package as a GGUF derivative of Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24. That original repo is a finetune of mistralai/Mistral-Nemo-Instruct-2407. The original Vikhr config records MistralForCausalLM geometry with 40 layers, 32 attention heads, 8 KV heads, 128 head dimension, 5120 hidden size, untied embeddings, and 1024000 max positions. The selected GGUF header records llama.cpp architecture llama but the same Mistral-Nemo-compatible geometry and the same 131074-token vocabulary." }, "architecture": { "canonical_architecture_id": "vikhr-nemo-12b-instruct", "max_context_tokens": 1024000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 12.24780288, "swept_params_b": 11.576704, "auxiliary_resident_params_b": 0.67109888, "resident_weight_gb": 13.02239104, "swept_weight_gb": 12.30146624, "auxiliary_resident_weight_gb": 0.7209248, "resident_parameter_scope": "selected GGUF linked file size and header tensor parameters for vikhr-nemo-12b-instruct-r-21-09-24-q8_0.gguf", "swept_parameter_scope": "ordinary text decode charges blk.* tensors, output_norm.weight, and output.weight from the selected Q8_0 GGUF artifact", "auxiliary_scope": "token_embd.weight plus GGUF metadata, tokenizer, header, and file overhead are resident in the selected artifact but not swept for ordinary text decode", "notes": "The selected Q8_0 linked file is 13.022391040 GB. GGUF tensor payloads total 13.014508800 GB, while GGUF metadata, tokenizer, header, and file alignment account for 0.007882240 GB. Because output.weight is stored separately, token_embd.weight is treated as resident-only input embedding traffic for ordinary text decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The selected GGUF header records 40 decoder layers, 8 KV heads, 128-dimensional key and value heads, no sliding-window metadata, and 1024000 context length. The serving profile charges ordinary full-context FP16 K/V cache streams for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the repo's single Q8_0 GGUF artifact. The selected file contains only text tensors and no mmproj, vision, audio, MTP, or draft tensors." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 1.0632430295938924, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q8-0-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor payload bytes for swept weight traffic. Dequantization, llama.cpp kernels, tokenizer processing, scheduler behavior, and compute overhead are outside Bounds Engine v1.", "notes": "The repo exposes a single Q8_0 GGUF file and HF API gguf.totalFileSize exactly matches that linked object. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "VlSav Vikhr-Nemo 12B Q8_0 GGUF API metadata", "url": "https://huggingface.co/api/models/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "downloads", "license", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "At commit 6a36ecf1d5586f8ce0d4a29d4c95ef2f3ad7f0e1, the API records a public non-gated GGUF repo with base_model Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24, Apache-2.0 license, English/Russian tags, region:us, GGUF architecture llama, context length 1024000, gguf.total 12247802880, gguf.totalFileSize 13022391040, and current downloads 199862. The repo contains one GGUF file, so gguf.totalFileSize selects the Q8_0 artifact." }, { "label": "VlSav Vikhr-Nemo 12B Q8_0 GGUF model card", "url": "https://huggingface.co/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "runtime_format", "selected_artifact" ], "notes": "The model card states this repo was converted to GGUF from Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24 using llama.cpp through GGUF-my-repo and shows llama.cpp CLI/server examples using vikhr-nemo-12b-instruct-r-21-09-24-q8_0.gguf." }, { "label": "Vikhr-Nemo 12B original API metadata", "url": "https://huggingface.co/api/models/Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24", "source_type": "model_card", "supports": [ "base_model_proof", "license", "logical_parameter_split" ], "notes": "At commit 7499757d9f41a2965b0e9db94c976e0982e292b8, the original repo is public, Apache-2.0 licensed, text-generation Transformers, and records base_model mistralai/Mistral-Nemo-Instruct-2407 plus BF16 safetensors parameters 12247802880." }, { "label": "Vikhr-Nemo 12B original config", "url": "https://huggingface.co/Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24/raw/7499757d9f41a2965b0e9db94c976e0982e292b8/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "embedding_layout", "max_context_tokens" ], "notes": "The config records MistralForCausalLM, model_type mistral, BF16 dtype, 40 layers, hidden size 5120, intermediate size 14336, 32 attention heads, 8 KV heads, 128 head dimension, max_position_embeddings 1024000, sliding_window null, rope_theta 1000000, tie_word_embeddings false, and vocab_size 131074. The _name_or_path and selected GGUF metadata still name R-05-09-24, but the public repo id and card are R-21-09-24." }, { "label": "Mistral Nemo Instruct base config comparison", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407/raw/main/config.json", "source_type": "config", "supports": [ "base_model_proof", "architecture", "kv_adapter" ], "notes": "Manual comparison found the same MistralForCausalLM layer, hidden, attention, KV, head-dimension, untied-embedding, and RoPE geometry as the original Vikhr config. The Vikhr repo extends max_position_embeddings to 1024000 and vocabulary size to 131074." }, { "label": "VlSav Vikhr-Nemo 12B Q8_0 GGUF linked-object size check", "url": "https://huggingface.co/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF/tree/6a36ecf1d5586f8ce0d4a29d4c95ef2f3ad7f0e1", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb" ], "notes": "The repo tree reports a single linked GGUF object, vikhr-nemo-12b-instruct-r-21-09-24-q8_0.gguf, with size 13022391040 bytes. This exactly matches API gguf.totalFileSize." }, { "label": "VlSav Vikhr-Nemo 12B Q8_0 GGUF range-read tensor index", "url": "https://huggingface.co/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF/resolve/6a36ecf1d5586f8ce0d4a29d4c95ef2f3ad7f0e1/vikhr-nemo-12b-instruct-r-21-09-24-q8_0.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 43 metadata entries and 363 tensors. The selected file is 13.022391040 GB, with tensor payloads starting at byte 7882240. Tensor payloads sum to 13.014508800 GB across 12.247802880B logical elements: token_embd.weight 0.713042560 GB, output.weight 0.713042560 GB, blk.* tensors 11.588403200 GB, and output_norm.weight 0.000020480 GB. Metadata/tokenizer/header/file overhead accounts for 0.007882240 GB. Stored tensor bytes split into Q8_0 13.012849920 GB and F32 0.001658880 GB. The header records general.architecture llama, base model mistralai/Mistral-Nemo-Instruct-2407, llama.block_count 40, context_length 1024000, embedding_length 5120, feed_forward_length 14336, attention.head_count 32, attention.head_count_kv 8, key/value length 128, rope.freq_base 1000000, and vocab_size 131074." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, original Vikhr API/config metadata, Mistral Nemo base config comparison, linked GGUF file size check, and direct selected Q8_0 GGUF tensor-index range read." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact selected Q8_0 GGUF file size, exact tensor spans, and resident-only input embedding plus GGUF overhead split." }, { "id": "vltx--vertalily-1-2-1b-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "VLTX/VertaLily-1.2-1B-GGUF", "title": "VertaLily 1.2 1B GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended VertaLily 1.2 1B Q4_K_M GGUF artifact.", "model_family": "lfm2-hybrid-gguf", "architecture": { "canonical_architecture_id": "lfm2-vltx-1-2b", "max_context_tokens": 128000, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 1.170340608, "swept_params_b": 1.170340608, "auxiliary_resident_params_b": 0, "resident_weight_gb": 0.730894368, "swept_weight_gb": 0.72850944, "auxiliary_resident_weight_gb": 0.002384928, "resident_parameter_scope": "selected VertaLily-1.2-1B-Q4_K_M-stable.gguf logical GGUF tensor elements plus GGUF metadata/header/file overhead", "swept_parameter_scope": "all tensor spans in the selected Q4_K_M GGUF artifact, including tied token_embd.weight output-projection traffic", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment/file overhead are resident but not swept as tensor traffic", "notes": "The selected Q4_K_M artifact has no separate output.weight tensor. LFM2 defaults to tied embeddings, so token_embd.weight is charged as ordinary decode output-projection traffic. Tensor spans sum to 0.728509440 GB; linked file size is 0.730894368 GB, leaving 0.002384928 GB of non-tensor GGUF overhead." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 6, "kv_heads": 8, "head_dim": 64, "kv_scalar_multiplier": 2, "notes": "The selected GGUF metadata and tensor names show attention projection tensors only on layers 2, 5, 8, 10, 12, and 14. These six layers use full-context K/V cache." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.00012288, "read_gb_per_output_token": 0.00012288, "state_formula": "10 short-conv layers * 2048 hidden channels * conv_L_cache 3 * 2 FP16 bytes", "notes": "The LFM2 implementation caches short-conv state through past_key_values.update_conv_state. For llama.cpp-style GGUF serving, this profile charges the short-conv state as activation-side FP16. Read traffic charges one full fixed-state read per generated token; compute and state writes remain outside Bounds Engine v1." } ], "notes": "The GGUF is LFM2-style hybrid state: six full-attention layers with context-growing K/V plus ten short-conv layers with fixed recurrent state." }, "notes": "The repo card brands the architecture as vltx, while the selected GGUF header records general.architecture lfm2. This profile uses the selected GGUF header and upstream LFM2 implementation for memory-side bounds." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6245142337229743, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-lfm2-hybrid-memory-bound", "dequantization_notes": "The memory-side bound charges exact GGUF tensor spans for ordinary swept weight traffic and the selected linked artifact size for residency. GGUF loader overhead, dequantization kernels, scheduler behavior, and state writes are outside Bounds Engine v1.", "notes": "The model card recommends the Q4_K_M artifact as the balanced general-use file. The HF API gguf.totalFileSize currently points to prop.PearlLily-unst.gguf, so this profile intentionally pins the card-recommended stable Q4_K_M linked object." }, "evidence": [ { "label": "VertaLily Hugging Face API metadata", "url": "https://huggingface.co/api/models/VLTX/VertaLily-1.2-1B-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "downloads", "license", "pipeline", "total_params_b", "max_context_tokens", "selected_artifact_mismatch" ], "notes": "The live API response at commit be658e2275a7f717e4648b9513ff5fe49bf68ff1 records a public non-gated Apache-2.0 text-generation GGUF repo with 152564 downloads, region:us, gguf.total 1170340608, gguf.architecture vltx, context_length 128000, and gguf.totalFileSize 695750848. The API totalFileSize matches prop.PearlLily-unst.gguf, not the card-recommended stable Q4_K_M artifact selected for this profile." }, { "label": "VertaLily model card", "url": "https://huggingface.co/VLTX/VertaLily-1.2-1B-GGUF/blob/be658e2275a7f717e4648b9513ff5fe49bf68ff1/README.md", "source_type": "model_card", "supports": [ "selected_artifact", "license", "pipeline", "serving", "model_family" ], "notes": "The pinned card identifies architecture vltx and lists Q3_K, Q4_K_M, and Q8_0 stable GGUF files. It describes VertaLily-1.2-1B-Q4_K_M-stable.gguf as the balanced sweet spot and ideal for most general use, and its Ollama example loads that Q4_K_M file." }, { "label": "VertaLily Q4_K_M linked object and GGUF header range-read", "url": "https://huggingface.co/VLTX/VertaLily-1.2-1B-GGUF/resolve/be658e2275a7f717e4648b9513ff5fe49bf68ff1/VertaLily-1.2-1B-Q4_K_M-stable.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "total_params_b", "weight_format", "layers", "kv_heads", "head_dim", "kv_adapter", "max_context_tokens" ], "notes": "HF linked-object metadata reports 730894368 bytes for VertaLily-1.2-1B-Q4_K_M-stable.gguf. A direct GGUF v3 range-read found 30 metadata entries and 148 tensors. Tensor spans sum to 0.728509440 GB across 1.170340608B logical elements; non-tensor metadata/header/file overhead is 0.002384928 GB. Payloads split into Q4_K 0.506068992 GB, Q6_K 0.221921280 GB, and F32 0.000519168 GB. The header records general.architecture lfm2, block_count 16, context_length 128000, embedding_length 2048, feed_forward_length 8192, attention.head_count 32, attention.head_count_kv [0,0,8,0,0,8,0,0,8,0,8,0,8,0,8,0], vocab_size 65536, and shortconv.l_cache 3. Tensor names show attention projections only on layers 2, 5, 8, 10, 12, and 14, with shortconv tensors on the other ten layers." }, { "label": "Transformers LFM2 configuration", "url": "https://github.com/huggingface/transformers/blob/5204b4fe36956e9214b9279f1e1be2fd5dd1d9f3/src/transformers/models/lfm2/configuration_lfm2.py", "source_type": "config", "supports": [ "model_family", "max_context_tokens", "tie_word_embeddings", "shortconv_state" ], "notes": "Manual review found Lfm2Config defaults tie_word_embeddings true, conv_L_cache 3, and builds layer_types from full_attn_idxs. This supports charging the selected GGUF's tied token_embd.weight as output-projection traffic and representing non-attention layers as fixed short-conv state." }, { "label": "Transformers LFM2 implementation", "url": "https://github.com/huggingface/transformers/blob/5204b4fe36956e9214b9279f1e1be2fd5dd1d9f3/src/transformers/models/lfm2/modeling_lfm2.py", "source_type": "manual_review", "supports": [ "kv_adapter", "attention_kv_geometry", "shortconv_state" ], "notes": "Manual review found Lfm2Attention updates K/V through past_key_values.update, while Lfm2ShortConv caches Bx state through update_conv_state with shape based on hidden channels and conv_L_cache. Lfm2DecoderLayer dispatches full_attention layers to attention and conv layers to Lfm2ShortConv." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, stable GGUF linked-object HEAD checks, a direct selected-GGUF header/tensor-index range read, and upstream Transformers LFM2 config/runtime review." }, "notes": "Use this profile for the stable Q4_K_M VertaLily artifact in ordinary text-decode bounds. Other GGUF files in this repo have different resident and traffic bytes and require separate selection." }, { "id": "weni--llama-guard-3-8b-awq", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "Weni/Llama-Guard-3-8B-AWQ", "title": "Weni Llama Guard 3 8B AWQ", "summary": "Audited memory-side bounds profile for the Weni AWQ INT4 Llama Guard 3 8B artifact.", "model_family": "llama3.1-guard-dense-awq", "base_model_proof": { "base_model": "meta-llama/Llama-Guard-3-8B", "relation": "quantized", "source": "Served AWQ config _name_or_path, visible gated base-model API metadata, and direct safetensors header review", "config_compatible": false, "notes": "The served config _name_or_path identifies a local snapshot of meta-llama/Llama-Guard-3-8B. The base API metadata is visible and records the same BF16 logical parameter count, but raw base config and tensor index access remain gated in this audit environment, so direct config compatibility with the base cannot be independently verified. This profile audits the served Weni AWQ artifact directly from its public config and tensor headers." }, "architecture": { "canonical_architecture_id": "llama-guard-3-8b", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 8.030261248, "swept_params_b": 7.504924672, "auxiliary_resident_params_b": 0.525336576, "resident_weight_gb": 5.727854592, "swept_weight_gb": 4.67718144, "auxiliary_resident_weight_gb": 1.050673152, "resident_parameter_scope": "logical Llama Guard 3 8B parameters represented by safetensors qweight plus F16 tensors", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and includes model.layers tensors, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token", "notes": "Bounds use exact stored bytes from safetensors headers because the AWQ package mixes packed I32 qweight/qzeros tensors with F16 scales and unquantized F16 embedding/head/norm tensors. Logical parameter counts are reconstructed from the served tensor layout: I32 qweight tensors unpack 8x to logical weights, while qzeros and scales are storage/runtime overhead rather than ordinary logical model weights. The reconstructed logical count matches the API safetensors total." }, "kv_adapter": { "kind": "full_context", "layers": 32, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 32 layers, 8 KV heads, hidden size 4096, 32 attention heads, Llama 3 RoPE scaling, and no sliding-window setting, so this profile charges full-context K and V streams for ordinary cached text decode." }, "notes": "Dense LlamaForCausalLM AWQ profile using the served Weni repo config and tensor headers. The profile does not rely on direct access to the gated Meta base config." }, "serving": { "weight_format": "int4", "weight_bytes_per_param": 0.7132837170679306, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-autoawq-awq-gemm-memory-bound", "dequantization_notes": "The memory-side bound charges stored AWQ packed weights, zero tensors, F16 scales, and unquantized F16 tensors from safetensors headers. AWQ dequantization, GEMM kernel differences, activation traffic, and compute overhead are outside Bounds Engine v1.", "notes": "The config records torch_dtype float16 and AWQ 4-bit GEMM quantization with group_size 128 and zero_point true. KV cache is charged at two bytes per scalar. weight_bytes_per_param summarizes resident stored bytes divided by reconstructed logical model parameters; exact resident/swept byte fields drive production bounds." }, "evidence": [ { "label": "Weni Llama Guard 3 8B AWQ API metadata", "url": "https://huggingface.co/api/models/Weni/Llama-Guard-3-8B-AWQ", "source_type": "model_card", "supports": [ "repo", "downloads", "serving", "commit_sha", "logical_parameter_count" ], "notes": "At commit 2157d713ed408803437e1a22f11d2c6dcb1034e2, the live API records a public non-gated safetensors repo with llama, 4-bit, awq, and region:us tags. Current downloads are 245818. The API safetensors block reports I32 6979321856, F16 1050939392, total 8030261248, matching the reconstructed logical parameter count from tensor headers." }, { "label": "Weni Llama Guard 3 8B AWQ served config", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/2157d713ed408803437e1a22f11d2c6dcb1034e2/config.json", "source_type": "config", "supports": [ "architecture", "kv_adapter", "serving", "quantization", "max_context_tokens", "base_model_proof" ], "notes": "The config records _name_or_path /root/.cache/huggingface/hub/models--meta-llama--Llama-Guard-3-8B/snapshots/5897b2ee0736e976b395039dea20a688a52bb31d, LlamaForCausalLM, torch_dtype float16, hidden_size 4096, intermediate_size 14336, 32 layers, 32 attention heads, 8 KV heads, max_position_embeddings 131072, tie_word_embeddings false, vocab_size 128256, rope_theta 500000, Llama 3 rope_scaling factor 8 with original_max_position_embeddings 8192, rms_norm_eps 1e-5, and AWQ GEMM 4-bit quantization with group_size 128 and zero_point true." }, { "label": "Weni Llama Guard 3 8B AWQ quant config", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/2157d713ed408803437e1a22f11d2c6dcb1034e2/quant_config.json", "source_type": "config", "supports": [ "serving", "quantization" ], "notes": "The quant_config records AWQ, zero_point true, q_group_size 128, w_bit 4, version gemm, and no modules_to_not_convert override." }, { "label": "Meta Llama Guard 3 8B base API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-Guard-3-8B", "source_type": "model_card", "supports": [ "base_model_proof", "logical_parameter_count", "license" ], "notes": "The current base API records gated manual access at commit 7327bd9f6efbbe6101dc6cc4736302b3cbb6e425, text-generation pipeline, Llama 3.1 license, region:us, base_model metadata for meta-llama/Llama-3.1-8B, and BF16 safetensors total 8030261248 parameters. Raw base config remains inaccessible in this audit environment, so the served AWQ config is the architecture source of truth." }, { "label": "Weni Llama Guard 3 8B AWQ safetensors index and shard headers", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/2157d713ed408803437e1a22f11d2c6dcb1034e2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "dtype_split", "storage_overhead", "embedding_layout" ], "notes": "The safetensors index records total_size 5727854592 bytes across two shards, matching direct range-read shard header spans. Headers contain 739 tensors totaling 5.727854592 GB: I32 3.516923904 GB and F16 2.210930688 GB. Stored suffix totals are qweight 3.489660928 GB, qzeros 0.027262976 GB, scales 0.109051904 GB, and F16 weight tensors 2.101878784 GB. model.embed_tokens.weight and lm_head.weight each have shape [128256, 4096] and contribute 0.525336576B parameters / 1.050673152 GB. Ordinary text swept traffic excluding only model.embed_tokens.weight totals 4.677181440 GB." }, { "label": "Weni Llama Guard 3 8B AWQ linked-object HEAD checks", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/tree/2157d713ed408803437e1a22f11d2c6dcb1034e2", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "commit_sha" ], "notes": "HEAD checks for both safetensors shards found linked sizes 4.677265296 GB and 1.050673280 GB. The linked file sizes include safetensors JSON header/container overhead; the index total_size and tensor data_offsets provide the resident tensor bytes used by the profile." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, served AWQ config, quant_config, gated base API metadata, safetensors index metadata, linked-object HEAD checks, direct shard header byte grouping, and local scrape row." }, "notes": "Use this profile for the Weni AWQ INT4 artifact only. The executable artifact is audited directly and does not turn the gated Meta BF16 base profile into an audited profile." }, { "id": "xiaomimimo--mimo-7b-base", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "XiaomiMiMo/MiMo-7B-Base", "title": "Xiaomi MiMo 7B Base BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 Xiaomi MiMo 7B Base repo.", "model_family": "mimo-qwen2-dense", "architecture": { "canonical_architecture_id": "mimo-7b-qwen2-dense-mtp1", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.833409536, "swept_params_b": 7.001337856, "auxiliary_resident_params_b": 0.83207168, "resident_weight_gb": 15.666819072, "swept_weight_gb": 14.002675712, "auxiliary_resident_weight_gb": 1.66414336, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and model.mtp_layers, and includes model.layers, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token; model.mtp_layers.0 is resident in this repo but is only used by MiMo MTP/speculative serving paths outside Bounds Engine v1 ordinary decode", "notes": "Range-read safetensors headers record 451 BF16 tensors totaling 7833409536 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. The custom MiMo model instantiates one MTP layer group, but ordinary Hugging Face MiMoForCausalLM inherits the Qwen2ForCausalLM forward path and does not sweep model.mtp_layers during standard target-model decode." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 36 layers, 8 KV heads, 128 head dimension, use_cache true, max_window_layers 32, sliding_window 32768, and use_sliding_window false. The custom modeling file imports and uses Qwen2Attention, which updates ordinary K/V past_key_values in the upstream Qwen2 path, so Bounds Engine v1 charges full-context BF16 K/V streams." }, "notes": "MiMoForCausalLM is a custom-code Qwen2-derived dense language model with one resident MTP layer group. This profile models ordinary text decode, not MiMo-MTP speculative serving." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. MiMo-MTP speculative decoding requires a separate paired/speculative adapter and is outside this ordinary decode profile." }, "evidence": [ { "label": "Xiaomi MiMo 7B Base API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-7B-Base", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "pipeline", "total_params_b", "weight_format", "commit_sha" ], "notes": "At commit c72df4586cb8bdeebd65f36929cd3385a6566fbe, the API records a public non-gated MIT text-generation Transformers repo with safetensors, custom_code, arxiv:2505.07608, eval-results, region:us, and 242871 downloads. The API safetensors block records BF16 7833409536 and total 7833409536." }, { "label": "Xiaomi MiMo 7B Base model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/README.md", "source_type": "model_card", "supports": [ "serving", "mtp_scope", "license" ], "notes": "The card describes MiMo-7B as a from-scratch reasoning model series and says the opened checkpoints include the base model. It states MIT licensing and points to the MiMo technical report." }, { "label": "Xiaomi MiMo 7B Base config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "mtp_scope", "serving" ], "notes": "The served config records MiMoForCausalLM, model_type mimo, BF16, hidden_size 4096, intermediate_size 11008, 36 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, max_window_layers 32, sliding_window 32768, use_sliding_window false, rope_theta 640000, tie_word_embeddings false, vocab_size 151680, attention_bias true, use_cache true, and num_nextn_predict_layers 1." }, { "label": "Xiaomi MiMo 7B Base custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/modeling_mimo.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "kv_adapter", "mtp_scope" ], "notes": "Manual review found MiMoModel extends Qwen2Model and appends self.mtp_layers after super().__init__(config). MiMoForCausalLM extends Qwen2ForCausalLM, sets self.model = MiMoModel(config), and creates a separate lm_head, but does not override forward. Therefore ordinary Transformers target-model decode uses the Qwen2 model/lm_head forward path and does not sweep model.mtp_layers unless an external MiMo-MTP serving path explicitly invokes it." }, { "label": "Xiaomi MiMo 7B Base safetensors index and shard headers", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "mtp_scope" ], "notes": "The index records total_size 15666819072 bytes across four shards. Range-read shard headers found 451 BF16 tensors totaling 7833409536 parameters / 15.666819072 GB, matching the index total and HF API safetensors total. Ordinary swept traffic is model.layers 12.760104960 GB plus model.norm.weight 0.000008192 GB plus lm_head.weight 1.242562560 GB, totaling 14.002675712 GB. Resident-only ordinary-decode tensors are model.embed_tokens.weight 1.242562560 GB and model.mtp_layers.0 tensors 0.421580800 GB, totaling 1.664143360 GB. Linked-object header/container overhead totals 51864 bytes outside tensor payloads." } ], "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, generation config, custom modeling code, safetensors index, and direct safetensors shard header range reads." }, "notes": "This profile is self-contained for ordinary target-model text decode. It intentionally does not model MiMo-MTP speculative decoding; that should become a separate adapter if Bounds Engine v1 grows explicit MTP/speculative support." }, { "id": "xiaomimimo--mimo-7b-rl", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "XiaomiMiMo/MiMo-7B-RL", "title": "Xiaomi MiMo 7B RL BF16", "summary": "Audited memory-side ordinary text-decode bounds profile for the BF16 Xiaomi MiMo 7B RL repo.", "model_family": "mimo-qwen2-dense", "architecture": { "canonical_architecture_id": "mimo-7b-qwen2-dense-mtp1", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 7.833409536, "swept_params_b": 7.001337856, "auxiliary_resident_params_b": 0.83207168, "resident_weight_gb": 15.666819072, "swept_weight_gb": 14.002675712, "auxiliary_resident_weight_gb": 1.66414336, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode excludes model.embed_tokens.weight input lookup and model.mtp_layers, and includes model.layers, model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.embed_tokens.weight is resident for token lookup but is not swept as a full matrix for each ordinary decode token; model.mtp_layers.0 is resident in this repo but is only used by MiMo MTP/speculative serving paths outside Bounds Engine v1 ordinary decode", "notes": "Range-read safetensors headers record 451 BF16 tensors totaling 7833409536 stored parameters. The config marks tie_word_embeddings false, and the checkpoint stores separate model.embed_tokens.weight and lm_head.weight tensors. The custom MiMo model instantiates one MTP layer group, but ordinary Hugging Face MiMoForCausalLM inherits the Qwen2ForCausalLM forward path and does not sweep model.mtp_layers during standard target-model decode." }, "kv_adapter": { "kind": "full_context", "layers": 36, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 36 layers, 8 KV heads, 128 head dimension, use_cache true, max_window_layers 36, sliding_window 32768, and use_sliding_window false. The custom modeling file imports and uses Qwen2Attention, which updates ordinary K/V past_key_values in the upstream Qwen2 path, so Bounds Engine v1 charges full-context BF16 K/V streams." }, "notes": "MiMoForCausalLM is a custom-code Qwen2-derived dense language model with one resident MTP layer group. This profile models ordinary text decode, not MiMo-MTP speculative serving." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The repo config records torch_dtype bfloat16, and range-read safetensors shard headers record only BF16 tensors. MiMo-MTP speculative decoding requires a separate paired/speculative adapter and is outside this ordinary decode profile." }, "evidence": [ { "label": "Xiaomi MiMo 7B RL API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-7B-RL", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "pipeline", "total_params_b", "weight_format", "commit_sha" ], "notes": "At commit 6299b5a2c45daf0c429285c92b8e61a5bd011c0d, the API records a public non-gated MIT text-generation Transformers repo with safetensors, custom_code, arxiv:2505.07608, eval-results, region:us, and 405357 downloads. The API safetensors block records BF16: 7833409536 and total: 7833409536." }, { "label": "Xiaomi MiMo 7B RL model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/README.md", "source_type": "model_card", "supports": [ "serving", "mtp_scope", "license" ], "notes": "The card describes MiMo-7B-RL as an RL model trained from the SFT model, documents one MTP layer for speculative decoding with about 90% acceptance rate, recommends a Xiaomi vLLM fork for MiMo-MTP, and also documents a vLLM registration path without loading MTP parameters. It states MIT licensing." }, { "label": "Xiaomi MiMo 7B RL config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "embedding_layout", "mtp_scope", "serving" ], "notes": "The served config records MiMoForCausalLM, model_type mimo, BF16, hidden_size 4096, intermediate_size 11008, 36 layers, 32 attention heads, 8 KV heads, head_dim 128, max_position_embeddings 32768, max_window_layers 36, sliding_window 32768, use_sliding_window false, rope_theta 640000, tie_word_embeddings false, vocab_size 151680, attention_bias true, use_cache true, and num_nextn_predict_layers 1." }, { "label": "Xiaomi MiMo 7B RL custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/modeling_mimo.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "kv_adapter", "mtp_scope" ], "notes": "Manual review found MiMoModel extends Qwen2Model and appends self.mtp_layers after super().__init__(config). MiMoForCausalLM extends Qwen2ForCausalLM, sets self.model = MiMoModel(config), and creates a separate lm_head, but does not override forward. Therefore ordinary Transformers target-model decode uses the Qwen2 model/lm_head forward path and does not sweep model.mtp_layers unless an external MiMo-MTP serving path explicitly invokes it." }, { "label": "Xiaomi MiMo 7B RL safetensors index and shard headers", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "resident_params_b", "swept_params_b", "embedding_layout", "mtp_scope" ], "notes": "The index records total_size 15666819072 bytes across four shards. Range-read shard headers found 451 BF16 tensors totaling 7833409536 parameters / 15.666819072 GB, matching the index total and HF API safetensors total. Ordinary swept traffic is model.layers 12.760104960 GB plus model.norm.weight 0.000008192 GB plus lm_head.weight 1.242562560 GB, totaling 14.002675712 GB. Resident-only ordinary-decode tensors are model.embed_tokens.weight 1.242562560 GB and model.mtp_layers.0 tensors 0.421580800 GB, totaling 1.664143360 GB. Linked-object HEAD checks resolved all four shards to 15.666870936 GB, leaving 51864 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, generation config, custom configuration/modeling code, safetensors index, direct safetensors shard header range reads, and linked-object HEAD checks." }, "notes": "This profile is self-contained for ordinary target-model text decode. It intentionally does not model MiMo-MTP speculative decoding; that should become a separate adapter if Bounds Engine v1 grows explicit MTP/speculative support." }, { "id": "xiaomimimo--mimo-v2-5-pro", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "XiaomiMiMo/MiMo-V2.5-Pro", "title": "Xiaomi MiMo V2.5 Pro FP8", "summary": "Audited memory-side ordinary text-decode bounds profile for the FP8 Xiaomi MiMo V2.5 Pro MoE language repo.", "model_family": "mimo-v2-hybrid-swa-moe", "architecture": { "canonical_architecture_id": "mimo-v2-5-pro-1023b-a41b-hybrid-swa-moe", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1033.369538304, "main_resident_weight_gb": 1029.031048704, "auxiliary_resident_weight_gb": 4.3384896, "fixed_weight_gb": 28.596352512, "routed_expert_weight_gb": 2.605298688, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding and MTP sidecar tensors", "auxiliary_scope": "model.embed_tokens.weight and model.mtp tensors are resident in the package but are not swept as full matrices for each ordinary target-model text decode token", "shared_expert_notes": "The served config records no shared experts. The single dense layer-0 MLP is included in fixed_weight_gb; the 69 MoE layers use eight routed experts per token.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 weights, BF16 head/embedding/MTP tensors, and F32 scale/bias tensors. Routed expert tensors are byte-uniform across all 384 expert indexes and 69 MoE layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 8, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Full-attention K cache. The served config and MiMoV2Attention code use num_key_value_heads 8 and head_dim 192 for non-SWA layers." }, { "kind": "full_context", "layers": 10, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Full-attention V cache. The served config records v_head_dim 128, so K and V are represented as separate scalar streams." }, { "kind": "sliding_window", "layers": 60, "kv_heads": 8, "head_dim": 192, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "SWA K cache. The runnable config and custom code use swa_num_key_value_heads 8, swa_head_dim 192, and a 128-token sliding window for SWA layers." }, { "kind": "sliding_window", "layers": 60, "kv_heads": 8, "head_dim": 128, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "SWA V cache. The runnable config and custom code use swa_v_head_dim 128." } ], "notes": "MiMo V2.5 Pro text decode is represented as BF16 K/V for ten full-context layers plus BF16 K/V for 60 128-token sliding-window layers. Bounds Engine v1 charges ordinary target-model decode only; MTP speculative sidecar execution is outside scope." }, "notes": "This profile models ordinary target-model text decode. MTP speculative decoding and any runtime-specific expert-parallel communication are outside Bounds Engine v1." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.009894817085528, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "sglang-vllm-fp8-hybrid-swa-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored safetensors bytes and BF16 KV bytes. FP8 dequantization, activation traffic, router compute, expert compute, MTP speculative sidecar execution, cache writes, and distributed all-to-all traffic are outside this memory-side bound.", "notes": "The config records dtype bfloat16 and FP8 e4m3 dynamic weight quantization with 128 by 128 blocks. No audited FP8 KV-cache setting is recorded in the served config, so KV is charged as BF16." }, "evidence": [ { "label": "Xiaomi MiMo V2.5 Pro API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5-Pro", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "total_params_b", "weight_format", "commit_sha" ], "notes": "At repo SHA f5953f1b993e90367f3d87fed59d099f2c527763, the API records a public non-gated MIT repo with safetensors, mimo_v2, text-generation, agent, long-context, code, conversational, custom_code, en, zh, eval-results, fp8, and region:us tags. Current downloads are 101692. The API safetensors block records F32 224695296, BF16 9450733440, F8_E4M3 1013569290240, and total 1023244718976 parameters." }, { "label": "Xiaomi MiMo V2.5 Pro model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/README.md", "source_type": "model_card", "supports": [ "model_family", "active_params_b", "max_context_tokens", "routed_experts", "routed_experts_per_token", "layer_pattern", "serving" ], "notes": "The card describes MiMo V2.5 Pro as an open-source MoE language model with 1.02T total parameters, 42B activated parameters, 1M context, FP8 weights, 70 layers, one dense layer plus 69 MoE layers, 10 full-attention layers, 60 SWA layers, 128 attention heads, 8 KV heads, QK/V head dimensions 192/128, 384 routed experts, 8 experts per token, 128-token SWA, and three MTP layers. The SGLang example uses a 1048576-token deployment context and speculative EAGLE/MTP options." }, { "label": "Xiaomi MiMo V2.5 Pro config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_adapter", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The served config records MiMoV2ForCausalLM, model_type mimo_v2, dtype bfloat16, FP8 e4m3 dynamic quantization, 70 layers, hidden size 6144, one dense MLP layer followed by MoE layers, moe_intermediate_size 2048, n_routed_experts 384, num_experts_per_tok 8, vocab_size 152576, tie_word_embeddings false, max_position_embeddings 1048576, hybrid_layer_pattern with 10 full layers and 60 SWA layers, full-layer num_key_value_heads 8, head_dim 192, v_head_dim 128, SWA num_key_value_heads 8, swa_head_dim 192, swa_v_head_dim 128, and sliding_window 128." }, { "label": "Xiaomi MiMo V2.5 Pro custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/modeling_mimo_v2.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "kv_adapter", "mtp_scope" ], "notes": "Manual review found MiMoV2Attention selects separate full and SWA KV geometries from the config, forms k_size as num_key_value_heads times head_dim, forms v_size as num_key_value_heads times v_head_dim, and updates past_key_values with those K/V tensors. MiMoV2DecoderLayer maps hybrid_layer_pattern 1 to sliding_window_attention and 0 to full_attention. MiMoV2ForCausalLM builds the target model and lm_head for ordinary decode, while _keys_to_ignore_on_load_unexpected includes model.mtp.*." }, { "label": "Xiaomi MiMo V2.5 Pro safetensors index and shard headers", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 34 indexed shards. Stored tensors sum to the index total_size, 1033.369538304 GB, across 159581 tensors: 1013.569290240 GB F8_E4M3, 18.901466880 GB BF16, and 0.898781184 GB F32. Ordinary-language resident tensors, defined as model layers/norm plus lm_head and routed experts but excluding input embedding and MTP tensors, sum to 1029.031048704 GB. Auxiliary resident tensors sum to 4.338489600 GB: input embedding 1.874853888 GB and MTP 2.463635712 GB. Fixed ordinary text traffic sums to 28.596352512 GB, and routed expert tensors sum to 1000.434696192 GB, exactly 2.605298688 GB per expert index. Header-derived logical ordinary active parameters are 40.959508608B for one token route." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-08", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, custom modeling code review, safetensors index metadata, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated catalog estimate, which treated the package as a flat 1-byte MoE with full-context KV for all layers. The audited profile separates FP8/BF16/F32 stored bytes, resident-only input/MTP tensors, routed expert traffic, and hybrid full-context/sliding-window KV." }, { "id": "xiaomimimo--mimo-v2-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "XiaomiMiMo/MiMo-V2.5", "title": "Xiaomi MiMo V2.5 FP8", "summary": "Audited memory-side ordinary text-decode bounds profile for the FP8 Xiaomi MiMo V2.5 omnimodal MoE repo.", "model_family": "mimo-v2-hybrid-swa-moe-omnimodal", "architecture": { "canonical_architecture_id": "mimo-v2-5-310b-a15b-hybrid-swa-moe", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 315.031102208, "main_resident_weight_gb": 310.612355968, "auxiliary_resident_weight_gb": 4.41874624, "fixed_weight_gb": 7.743236992, "routed_expert_weight_gb": 1.183082496, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.layers, model.norm, and lm_head, excluding resident-only input embedding, visual encoder, audio encoder, and MTP sidecar tensors", "auxiliary_scope": "model.embed_tokens.weight, visual tensors, audio_encoder tensors, speech_embeddings, and model.mtp tensors are resident in the package but are not swept as full matrices for each ordinary target-model text decode token", "shared_expert_notes": "The served config records no shared experts. The single dense layer-0 MLP is included in fixed_weight_gb; the 47 MoE layers use eight routed experts per token.", "notes": "Header-derived stored bytes are used because the package mixes F8_E4M3 weights, BF16 multimodal/embedding/head tensors, and F32 scale/bias tensors. Routed expert tensors are byte-uniform across all 256 expert indexes and 47 MoE layers." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 9, "kv_heads": 4, "head_dim": 192, "kv_scalar_multiplier": 1, "notes": "Full-attention K cache. The served config and MiMoV2Attention code use num_key_value_heads 4 and head_dim 192 for non-SWA layers." }, { "kind": "full_context", "layers": 9, "kv_heads": 4, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "Full-attention V cache. The served config records v_head_dim 128, so K and V are represented as separate scalar streams." }, { "kind": "sliding_window", "layers": 39, "kv_heads": 8, "head_dim": 192, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "SWA K cache. The runnable config and custom code use swa_num_key_value_heads 8, swa_head_dim 192, and a 128-token sliding window for SWA layers." }, { "kind": "sliding_window", "layers": 39, "kv_heads": 8, "head_dim": 128, "window_tokens": 128, "kv_scalar_multiplier": 1, "notes": "SWA V cache. The runnable config and custom code use swa_v_head_dim 128." } ], "notes": "MiMo V2.5 text decode is represented as BF16 K/V for nine full-context layers plus BF16 K/V for 39 128-token sliding-window layers. The model card's KV-head table appears reversed relative to the served config and custom code; this profile follows the runnable artifact." }, "notes": "This profile models ordinary target-model text decode after any multimodal prefill. Vision, video, audio preprocessing, multimodal prefill, and MTP speculative decoding are outside Bounds Engine v1." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.013694993677742, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "sglang-vllm-fp8-hybrid-swa-moe-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored safetensors bytes and BF16 KV bytes. FP8 dequantization, activation traffic, router compute, expert compute, multimodal encoder compute, MTP speculative sidecar execution, and cache writes are outside this memory-side bound.", "notes": "The config records dtype bfloat16 and FP8 e4m3 dynamic weight quantization with 128 by 128 blocks. No audited FP8 KV-cache setting is recorded in the served config, so KV is charged as BF16." }, "evidence": [ { "label": "Xiaomi MiMo V2.5 API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5", "source_type": "model_card", "supports": [ "repo", "license", "downloads", "total_params_b", "weight_format", "commit_sha" ], "notes": "At repo SHA 2fd4f899a491de2fb0beeafe32b5d700b251f593, the API records a public non-gated MIT repo with safetensors, custom_code, multimodal, vision-language, audio, agent, video-understanding, long-context, fp8, and region:us tags. Current downloads are 219382. The API safetensors block records F32 68012416, BF16 4052024960, F8_E4M3 306655002624, and total 310775040000 parameters." }, { "label": "Xiaomi MiMo V2.5 model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/README.md", "source_type": "model_card", "supports": [ "model_family", "active_params_b", "max_context_tokens", "routed_experts", "routed_experts_per_token", "layer_pattern", "modalities", "serving" ], "notes": "The card describes MiMo V2.5 as a native omnimodal sparse MoE with 310B total and 15B activated parameters, 48 language layers, one dense layer plus 47 MoE layers, 9 full-attention layers, 39 SWA layers, 256 routed experts, 8 experts per token, 128-token SWA, 1M context, a 729M-parameter vision encoder, a 261M-parameter audio encoder, and three MTP layers. The SGLang example uses a 262144-token deployment context, but the served config records 1048576 max positions." }, { "label": "Xiaomi MiMo V2.5 config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_adapter", "routed_experts", "routed_experts_per_token", "max_context_tokens", "serving", "auxiliary_resident_scope" ], "notes": "The served config records MiMoV2ForCausalLM, model_type mimo_v2, dtype bfloat16, FP8 e4m3 dynamic quantization, 48 layers, hidden size 4096, one dense MLP layer followed by MoE layers, moe_intermediate_size 2048, n_routed_experts 256, num_experts_per_tok 8, vocab_size 152576, tie_word_embeddings false, max_position_embeddings 1048576, hybrid_layer_pattern with 9 full layers and 39 SWA layers, full-layer num_key_value_heads 4, head_dim 192, v_head_dim 128, SWA num_key_value_heads 8, swa_head_dim 192, swa_v_head_dim 128, and sliding_window 128." }, { "label": "Xiaomi MiMo V2.5 custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/modeling_mimo_v2.py", "source_type": "manual_review", "supports": [ "ordinary_decode_scope", "kv_adapter", "mtp_scope", "modal_scope" ], "notes": "Manual review found MiMoV2Attention selects separate full and SWA KV geometries from the config, forms k_size as num_key_value_heads times head_dim, forms v_size as num_key_value_heads times v_head_dim, and updates past_key_values with those K/V tensors. MiMoV2DecoderLayer maps hybrid_layer_pattern 1 to sliding_window_attention and 0 to full_attention. MiMoV2ForCausalLM builds visual and audio modules but only invokes them when modal inputs are present, and _keys_to_ignore_on_load_unexpected includes model.mtp.*." }, { "label": "Xiaomi MiMo V2.5 safetensors index and shard headers", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "Safetensors headers were range-read across all 17 indexed shards. Stored tensors sum to the index total_size, 315.031102208 GB, across 73081 tensors: 306.655002624 GB F8_E4M3, 8.104049920 GB BF16, and 0.272049664 GB F32. Ordinary-language resident tensors, defined as model layers/norm plus lm_head and routed experts but excluding input embedding, visual, audio, and MTP tensors, sum to 310.612355968 GB. Auxiliary resident tensors sum to 4.418746240 GB: input embedding 1.249902592 GB, visual 1.457188864 GB, audio 0.522254336 GB, and MTP 1.189400448 GB. Fixed ordinary text traffic sums to 7.743236992 GB, and routed expert tensors sum to 302.869118976 GB, exactly 1.183082496 GB per expert index." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, pinned model card, pinned served config, custom configuration/modeling code review, safetensors index metadata, and direct safetensors shard-header range reads." }, "notes": "This profile replaces the generated catalog estimate, which treated the package as a flat 1-byte MoE with full-context KV for all layers. The audited profile separates FP8/BF16/F32 stored bytes, resident-only multimodal/input/MTP tensors, routed expert traffic, and hybrid full-context/sliding-window KV." }, { "id": "ykarout--qwen3-6-35b-a3b-nvfp4", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "ykarout/Qwen3.6-35B-A3B-NVFP4", "title": "ykarout Qwen3.6 35B A3B NVFP4", "summary": "Audited memory-side ordinary text-decode bounds profile for the ykarout ModelOpt NVFP4 Qwen3.6 35B A3B artifact.", "model_family": "qwen3.6-moe-multimodal", "base_model_proof": { "base_model": "Qwen/Qwen3.6-35B-A3B", "relation": "quantized", "source": "Hugging Face base_model metadata, served config comparison, hf_quant_config review, and direct safetensors header grouping", "config_compatible": true, "notes": "The repo metadata identifies Qwen/Qwen3.6-35B-A3B as the quantized base. Manual comparison with the current base config found matching checked architecture fields: Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, 40 text layers, full_attention_interval 4, hidden size 2048, 16 attention heads, 2 KV heads, 256 full-attention head dimension, DeltaNet state geometry, 256 experts, 8 routed experts per token, shared_expert_intermediate_size 512, untied embeddings, 262144 max positions, and the resident 27-layer vision config." }, "architecture": { "canonical_architecture_id": "qwen3-6-35b-a3b", "max_context_tokens": 262144, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 23.909370592, "main_resident_weight_gb": 21.999109376, "auxiliary_resident_weight_gb": 1.910261216, "fixed_weight_gb": 3.879593216, "routed_expert_weight_gb": 0.07077936, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_nvfp4_fp8_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model and lm_head, excluding resident-only input embedding and visual tensors", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for the multimodal package but are not swept for each ordinary text decode token", "shared_expert_notes": "The model card states 8 routed plus 1 shared expert, and the config records shared_expert_intermediate_size 512. Shared expert and shared expert gate tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used instead of the rounded 35B/3B model-card parameters. This ModelOpt artifact targets Linear modules but excludes visual modules, lm_head, every linear_attn module, every shared expert/shared_expert_gate, and every full-attention self_attn module. Routed expert tensors are byte-uniform across all 256 expert indexes." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 10, "kv_heads": 2, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "The language stack has 40 layers with every fourth layer using full attention, giving 10 full-context attention layers." }, { "kind": "recurrent_state", "alloc_gb_per_session": 0.06488064, "read_gb_per_output_token": 0.06488064, "state_formula": "30 linear_attention layers * ((32 repeated value heads * 128 key dim * 128 value dim * 4 F32 bytes) + ((16 key heads * 128 key dim * 2 + 32 value heads * 128 value dim) * 4 conv kernel * 2 BF16 bytes))", "notes": "The qwen3_5 implementation caches conv_state plus recurrent_state for linear-attention layers. ModelOpt quantizes weights and requests FP8 KV for full-attention layers, but it does not remove the fixed DeltaNet runtime state. Read traffic charges one full fixed-state read per generated token; compute and write traffic remain outside Bounds Engine v1." } ], "notes": "Hybrid Qwen3.6 text decode is represented as FP8 full-context K/V for full-attention layers plus fixed DeltaNet recurrent state for linear-attention layers." }, "notes": "Qwen3_5MoeForConditionalGeneration includes a vision tower. This profile models ordinary text decode through the language model and output head after any multimodal prefill, with visual tensors separated from per-token decode traffic." }, "serving": { "weight_format": "nvfp4", "weight_bytes_per_param": 0.5, "kv_store_format": "fp8", "kv_store_bytes_per_scalar": 1, "kv_read_format": "fp8", "kv_read_bytes_per_scalar": 1, "runtime_format": "vllm-modelopt-nvfp4-fp8-kv-text-decode-memory-bound", "dequantization_notes": "Bounds Engine v1 charges stored ModelOpt NVFP4 U8 weight payloads, F8 scale/side tensors, BF16 unquantized tensors, F32 scalar scale tensors, FP8 full-attention KV, and BF16/F32 DeltaNet state bytes. Activation traffic, NVFP4 dequantization, router/expert compute, multimodal encoder compute, and cache writes are outside this memory-side bound.", "notes": "The model card title and hf_quant_config record NVFP4 weights with FP8 KV cache. The config does not include a compressed-tensors kv_cache_scheme, so the FP8 KV assumption comes from the ModelOpt hf_quant_config and model card." }, "evidence": [ { "label": "ykarout Qwen3.6 35B A3B NVFP4 API metadata", "url": "https://huggingface.co/api/models/ykarout/Qwen3.6-35B-A3B-NVFP4", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "revision", "downloads", "weight_format", "serving" ], "notes": "At commit 5093c2a1d4d263007b563753438caf75dadaec48, the live API records a public non-gated Apache-2.0 image-text-to-text repo with qwen3_5_moe, NVFP4, 4-bit, FP4, Blackwell, modelopt, endpoints_compatible, and region:us tags. Current downloads are 202525. The API safetensors block records BF16 2894927216, F8_E4M3 2013265920, U8 16106127360, and total 35951800000." }, { "label": "ykarout Qwen3.6 35B A3B NVFP4 model card", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "total_params_b", "active_params_b", "layers", "kv_heads", "head_dim", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving" ], "notes": "The pinned card describes this repo as Qwen3.6-35B-A3B NVFP4 experts-only quantization with FP8 KV cache using NVIDIA Model Optimizer. The model overview states 35B total and 3B activated parameters, 40 layers, hidden size 2048, 10 repetitions of 3 Gated DeltaNet layers followed by 1 Gated Attention layer, 32 V heads and 16 QK heads for Gated DeltaNet, 16 Q heads and 2 KV heads for Gated Attention, head dimension 256, 256 experts, 8 routed plus 1 shared expert, 512 expert intermediate dimension, 248320 padded vocabulary, MTP training, and 262144 native context." }, { "label": "ykarout Qwen3.6 35B A3B NVFP4 config", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "linear_attention_state", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "quantization_ignore_scope" ], "notes": "The served config records Qwen3_5MoeForConditionalGeneration, qwen3_5_moe_text, untied embeddings, BF16 text dtype, 40 text layers, full_attention_interval 4, 10 full-attention layers, 30 linear-attention layers, 2 KV heads, 256 full-attention head dimension, 32 linear value heads, 16 linear key heads, 128 linear key/value dimensions, 4-token linear convolution kernel, mamba_ssm_dtype float32, 256 experts, 8 experts per token, shared_expert_intermediate_size 512, 262144 max position embeddings, and a resident vision_config." }, { "label": "ykarout Qwen3.6 35B A3B NVFP4 ModelOpt quantization config", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/hf_quant_config.json", "source_type": "config", "supports": [ "weight_format", "kv_store_format", "kv_read_format", "quantization_ignore_scope" ], "notes": "The pinned hf_quant_config records producer modelopt, quant_algo NVFP4, kv_cache_quant_algo FP8, group_size 16, and exclude_modules covering lm_head, model.visual*, all linear_attn modules, all shared_expert/shared_expert_gate modules, and full-attention self_attn modules." }, { "label": "Qwen3.6 35B A3B base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json", "source_type": "config", "supports": [ "base_model_proof", "layers", "kv_heads", "head_dim", "linear_attention_state", "max_context_tokens" ], "notes": "Manual comparison against the current base config at commit 995ad96eacd98c81ed38be0c5b274b04031597b0 found matching checked top-level, text_config, and vision_config geometry fields. The ykarout artifact adds ModelOpt quantization metadata while preserving the base architecture." }, { "label": "ykarout Qwen3.6 35B A3B NVFP4 safetensors index and shard headers", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "dtype_split" ], "notes": "The index records total_size 23.909370592 GB across three shards and total_parameters 35951800000. Direct range-read safetensors headers found 93106 tensors totaling the same 23.909370592 GB: F32 0.000122880 GB, BF16 5.789854432 GB, F8_E4M3 2.013265920 GB, and U8 16.106127360 GB. Ordinary text resident tensors, defined as model.language_model excluding embed_tokens plus lm_head, sum to 21.999109376 GB. Auxiliary resident tensors, defined as model.visual plus model.language_model.embed_tokens.weight, sum to 1.910261216 GB. Routed expert tensors sum to 18.119516160 GB and divide exactly into 256 uniform expert indexes of 0.070779360 GB. Fixed ordinary text traffic, including lm_head, DeltaNet weights, full-attention weights, shared experts, gates, and norms, sums to 3.879593216 GB." }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py", "source_type": "manual_review", "supports": [ "linear_attention_state", "kv_adapter" ], "notes": "Manual review at Transformers commit 36778030cba4bccfcb15366d7ad860ca30f47417 found Qwen3_5GatedDeltaNet caches conv_state and recurrent_state. The implementation defines conv_dim as key_dim * 2 + value_dim, repeats key/query heads to the value-head count before the recurrent kernel, and creates last_recurrent_state with shape [batch, num_heads, key_head_dim, value_head_dim]." } ], "review": { "reviewed_by": "Bob ", "reviewed_at": "2026-07-06", "notes": "Manual one-model audit from pinned ykarout API/card/config/quant-config/index evidence, current Qwen base config comparison, direct safetensors shard-header grouping, and the audited Qwen3.5 runtime state adapter." }, "notes": "This profile supersedes the generated half-byte estimate. It uses exact stored ModelOpt tensor bytes, charges FP8 full-attention KV plus fixed DeltaNet state, and keeps vision/input embedding tensors resident-only for ordinary text decode." }, { "id": "yuxinlu1--gemma-4-12b-agentic-fable5-composer2-5-v2-3-5x-tau2-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF", "title": "Yuxinlu Gemma 4 12B Agentic GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended Q4_K_M GGUF artifact of the Gemma 4 12B agentic v2 fine-tune.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "google/gemma-4-12B-it", "relation": "quantized", "source": "Hugging Face model card metadata, Google base config, direct GGUF header metadata, and missing repo-config check", "config_compatible": true, "notes": "The GGUF repo card identifies google/gemma-4-12B-it as the base model. The GGUF repo does not ship config.json, so this profile uses the immutable Google Gemma 4 12B IT config for high-level architecture fields and the selected GGUF header for exact tensor, context, and KV metadata. The selected GGUF header records the same Gemma 4 12B unified text geometry." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.381381664, "swept_weight_gb": 7.365559808, "auxiliary_resident_weight_gb": 0.015821856, "resident_parameter_scope": "card-recommended GGUF linked file size and selected Q4_K_M header tensor spans", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding outside the tensor span are resident in the selected artifact file but not swept as model tensors", "notes": "The profile targets gemma4-v2-Q4_K_M.gguf because the model card calls Q4_K_M the recommended sweet spot and uses it in the llama.cpp example command. The HF API gguf.totalFileSize points at gemma4-v2-Q3_K_M.gguf, so this profile records that evidence but does not use Q3_K_M as the repo-level ordinary-serving default. Header tensor spans total 7.365559808 GB, while the linked file size is 7.381381664 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected normal-use Q4_K_M GGUF artifact. It does not model speculative decoding from the separate MTP draft sidecars." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6199012632480672, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, and speculative decoding are outside Bounds Engine v1.", "notes": "The selected normal text artifact uses a mixed Q4_K_M layout: typed tensor payloads split into 5.246484480 GB Q4_K, 2.115993600 GB Q6_K, and 0.003080384 GB F32, plus 0.000001344 GB of tensor-alignment padding in the swept tensor span. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Yuxinlu Gemma 4 12B Agentic GGUF HF API metadata", "url": "https://huggingface.co/api/models/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 190a31365a6b80a692349be34ccdac730cad4fe4 records a public Apache-2.0 text-generation GGUF repo with base_model google/gemma-4-12B-it, region:us, 355871 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 6087086624. API totalFileSize matches gemma4-v2-Q3_K_M.gguf, while the model card recommends Q4_K_M for normal use." }, { "label": "Yuxinlu Gemma 4 12B Agentic GGUF model card", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF/raw/190a31365a6b80a692349be34ccdac730cad4fe4/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "mtp_sidecars" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model google/gemma-4-12B-it, coding/agentic specialization, a quant table that calls Q4_K_M the recommended sweet spot, and a llama.cpp command using gemma4-v2-Q4_K_M.gguf. It also documents separate MTP draft sidecars for speculative decoding." }, { "label": "Yuxinlu Gemma 4 12B Agentic missing repo config check", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF/raw/190a31365a6b80a692349be34ccdac730cad4fe4/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "The pinned repo-local config.json URL returns 404. The profile therefore does not assume an unserved local config and instead uses the selected GGUF header plus the Google base config." }, { "label": "Google Gemma 4 12B IT config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The immutable config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "Yuxinlu Gemma 4 12B Agentic GGUF linked-object HEAD checks", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF/tree/190a31365a6b80a692349be34ccdac730cad4fe4", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope", "mtp_sidecars" ], "notes": "HEAD checks found gemma4-v2-Q3_K_M.gguf 6.087086624 GB, Q4_K_M 7.381381664 GB, Q6_K 9.786020384 GB, and Q8_0 12.669645344 GB. Separate MTP sidecars are BF16 0.861520128 GB, F16 0.861520128 GB, and Q8_0 0.465109248 GB. The Q3_K_M size matches API gguf.totalFileSize, but Q4_K_M is the card-recommended normal-use artifact selected by this profile." }, { "label": "Yuxinlu Gemma 4 12B Agentic Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF/resolve/190a31365a6b80a692349be34ccdac730cad4fe4/gemma4-v2-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 43 metadata entries and 667 tensors. The selected file is 7.381381664 GB, with tensor payloads starting at byte 15821856. Tensor spans total 7.365559808 GB across 11907350576 logical elements: token_embd.weight 0.825753600 GB, blk.* tensors 6.539788480 GB, output_norm.weight 0.000015360 GB, rope_freqs.weight 0.000001024 GB, and 0.000001344 GB tensor-alignment padding. Typed tensor payloads split into Q4_K 5.246484480 GB, Q6_K 2.115993600 GB, and F32 0.003080384 GB. Metadata/tokenizer/header/file overhead outside the tensor span accounts for 0.015821856 GB. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from live HF API metadata, pinned model card, missing repo-local config check, immutable Google Gemma 4 12B IT config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the card-recommended Q4_K_M GGUF text artifact. Do not infer another quantization or speculative/MTP sidecar residency unless the workload explicitly selects those GGUF files." }, { "id": "yuxinlu1--gemma-4-12b-coder-fable5-composer2-5-v1-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF", "title": "Yuxinlu Gemma 4 12B Coder GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended Q4_K_M GGUF artifact of the Gemma 4 12B coder fine-tune.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1", "relation": "quantized", "source": "Hugging Face model card metadata, linked full-precision fine-tune config, Google base config comparison, and direct GGUF header metadata", "config_compatible": true, "notes": "The GGUF repo card identifies google/gemma-4-12B-it as the training base and links the full-precision yuxinlu1 safetensors fine-tune. The safetensors fine-tune config and Google base config have no differences in the text geometry used by this profile. The selected GGUF header records the same Gemma 4 12B text geometry." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 262144, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.381381664, "swept_weight_gb": 7.365559808, "auxiliary_resident_weight_gb": 0.015821856, "resident_parameter_scope": "card-recommended GGUF linked file size and selected Q4_K_M header tensor shapes", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors", "notes": "The profile targets gemma4-coding-Q4_K_M.gguf because the model card calls Q4_K_M the sweet spot and uses it in the llama.cpp example command. The HF API gguf.totalFileSize points at gemma4-coding-Q2_K.gguf, so this profile records that evidence but does not use Q2_K as the repo-level ordinary-serving default. Header tensor spans total 7.365559808 GB, while the linked file size is 7.381381664 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision, audio, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package recommends q8_0 KV in its example command, but Bounds Engine v1 uses FP16-equivalent KV byte traffic for llama.cpp-style GGUF serving profiles." }, "notes": "This profile models ordinary text decode for the selected normal-use Q4_K_M GGUF artifact. It does not model multimodal projector execution or speculative decoding sidecars." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.6199012632480672, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected normal text artifact uses a mixed Q4_K_M layout: tensor spans split into 5.24648448 GB Q4_K, 2.1159936 GB Q6_K, and 0.003081728 GB F32. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected linked file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "Yuxinlu Gemma 4 12B Coder GGUF HF API metadata", "url": "https://huggingface.co/api/models/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "total_params_b", "max_context_tokens", "selected_artifact" ], "notes": "The live HF API response at commit 1380be1796e559fca96b4107599285cab3ddbb92 records a public Apache-2.0 text-generation GGUF repo with base_model google/gemma-4-12B-it, region:us, 641260 downloads, GGUF architecture gemma4, 262144 context length, gguf.total 11907350576, and gguf.totalFileSize 4830147104. API totalFileSize matches gemma4-coding-Q2_K.gguf, while the model card recommends Q4_K_M for normal use." }, { "label": "Yuxinlu Gemma 4 12B Coder GGUF model card", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "selected_artifact", "max_context_tokens" ], "notes": "The card records Apache-2.0 licensing, text-generation packaging, base_model google/gemma-4-12B-it, a linked full-precision safetensors master, a corrected 256K context window, and a quant table that calls Q4_K_M the sweet spot. The llama.cpp example command uses gemma4-coding-Q4_K_M.gguf." }, { "label": "Yuxinlu Gemma 4 12B Coder safetensors config", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1/raw/237e5d8e72fb70e2c4d5df2676bc95308aad279c/config.json", "source_type": "config", "supports": [ "base_model_proof", "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The full-precision fine-tune config records Gemma4UnifiedForConditionalGeneration, tied embeddings, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, and 262144 max position embeddings." }, { "label": "Google Gemma 4 12B IT config comparison", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json", "source_type": "config", "supports": [ "base_model_proof", "kv_adapter" ], "notes": "Manual comparison found no differences between the full-precision fine-tune config and the Google Gemma 4 12B IT config in the text, serving, context, layer-type, and KV geometry fields used by this profile." }, { "label": "Yuxinlu Gemma 4 12B Coder GGUF linked-object HEAD checks", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/tree/1380be1796e559fca96b4107599285cab3ddbb92", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma4-coding-Q2_K.gguf 4.830147104 GB, Q3_K_M 6.087086624 GB, Q4_K_M 7.381381664 GB, Q6_K 9.786020384 GB, and Q8_0 12.669645344 GB. The Q2_K size matches API gguf.totalFileSize, but Q4_K_M is the card-recommended normal-use artifact selected by this profile." }, { "label": "Yuxinlu Gemma 4 12B Coder Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/resolve/1380be1796e559fca96b4107599285cab3ddbb92/gemma4-coding-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the GGUF v3 header found 43 metadata entries and 667 tensors. The selected file is 7.381381664 GB, with tensor payloads starting at byte 15821856. Tensor spans total 7.365559808 GB across 11907350576 logical elements: token_embd.weight 0.8257536 GB, blk.* tensors 6.539789824 GB, output_norm.weight 0.00001536 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_K 5.24648448 GB, Q6_K 2.1159936 GB, and F32 0.003081728 GB. Metadata/tokenizer/header/file overhead accounts for 0.015821856 GB. The header records gemma4.block_count 48, context_length 262144, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512 for global layers, key/value length 256 for sliding layers, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from live HF API metadata, model card, full-precision fine-tune config, Google base config comparison, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the card-recommended Q4_K_M GGUF text artifact. Do not infer another quantization unless the workload explicitly selects a different GGUF file." }, { "id": "zaakirio--gemma-4-12b-it-uncensored-gguf", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zaakirio/gemma-4-12b-it-uncensored-GGUF", "title": "zaakirio Gemma 4 12B IT Uncensored GGUF Q4_K_M", "summary": "Audited memory-side text-decode bounds profile for the card-recommended Q4_K_M GGUF artifact of the decensored Gemma 4 12B IT derivative.", "model_family": "gemma4-dense-unified-multimodal", "base_model_proof": { "base_model": "zaakirio/gemma-4-12b-it-uncensored", "relation": "quantized", "source": "Hugging Face model card/API metadata, source-model config, Google base provenance, and selected GGUF header metadata", "config_compatible": true, "notes": "The repo card and API metadata identify this package as a GGUF derivative of zaakirio/gemma-4-12b-it-uncensored. The card says that source model is a Heretic-abliterated derivative of google/gemma-4-12B-it and that ablation touches only the language weights. The selected Q4_K_M GGUF header and the source-model config agree on the Gemma 4 12B text geometry used by this profile: 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, tied embeddings, and 131072 max positions." }, "architecture": { "canonical_architecture_id": "gemma-4-12b-unified", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 11.907350576, "swept_params_b": 11.907350576, "auxiliary_resident_params_b": 0, "resident_weight_gb": 7.38138176, "swept_weight_gb": 7.365559808, "auxiliary_resident_weight_gb": 0.015821952, "resident_parameter_scope": "selected GGUF linked file size for gemma-4-12b-it-uncensored-Q4_K_M.gguf", "swept_parameter_scope": "ordinary text decode charges all GGUF tensor spans in the selected main text artifact; token_embd.weight is charged as the tied output projection because no output.weight tensor is stored", "auxiliary_scope": "GGUF metadata, tokenizer, header, and alignment padding are resident in the selected artifact file but not swept as model tensors; mmproj and MTP sidecar GGUF files are not included unless explicitly loaded for another workload", "notes": "The profile targets the Q4_K_M GGUF file selected by the model card recommendation and llama.cpp examples. The live API gguf.totalFileSize points to the f16 artifact, so this profile intentionally records that mismatch and pins the card-recommended Q4_K_M file. Header tensor spans total 7.365559808 GB, while the linked file size is 7.381381760 GB. The main GGUF contains only text tensors: rope_freqs, token_embd, blk.*, and output_norm. It has no mmproj, vision/audio projector, MTP, or output.weight tensor." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 8, "kv_heads": 1, "head_dim": 512, "kv_scalar_multiplier": 1, "notes": "Full-attention layers are 5, 11, 17, 23, 29, 35, 41, and 47. The config and GGUF metadata record one global KV head and 512 global key/value length; official Gemma4Attention uses the K=V path only for non-sliding full-attention layers, so this component charges one effective KV stream." }, { "kind": "sliding_window", "layers": 40, "kv_heads": 8, "head_dim": 256, "window_tokens": 1024, "kv_scalar_multiplier": 2, "notes": "The remaining 40 layers use 1024-token local sliding-window attention with eight KV heads and separate K/V streams." } ], "notes": "Hybrid local/global text attention is represented as an explicit layered KV sum. The selected GGUF package does not declare a required quantized KV cache format, so the serving profile uses FP16 KV for llama.cpp-style GGUF serving." }, "notes": "This profile models ordinary text decode for the selected main Q4_K_M GGUF artifact after any multimodal prefill. The separate BF16 mmproj sidecar and optional MTP drafter sidecar require separate workload profiles if loaded." }, "serving": { "weight_format": "gguf_quantized", "weight_bytes_per_param": 0.619901271310314, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "llama.cpp-gguf-q4-k-m-memory-bound", "dequantization_notes": "The memory-side bound charges exact selected GGUF file bytes for residency and exact GGUF tensor spans for swept weight traffic. Dequantization, kernels, scheduler behavior, multimodal projector execution, MTP drafter execution, and multimodal prefill traffic are outside Bounds Engine v1.", "notes": "The selected artifact is the card-recommended Q4_K_M GGUF. Explicit resident and swept byte fields are authoritative; weight_bytes_per_param summarizes selected file bytes divided by GGUF logical tensor elements." }, "evidence": [ { "label": "zaakirio Gemma 4 12B IT Uncensored GGUF API metadata", "url": "https://huggingface.co/api/models/zaakirio/gemma-4-12b-it-uncensored-GGUF", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "selected_artifact_mismatch", "downloads", "total_params_b", "max_context_tokens" ], "notes": "The live HF API response at commit 32880562ac43cb589a85afb864309fdcaf486fae records base_model zaakirio/gemma-4-12b-it-uncensored, Gemma license metadata, image-text-to-text pipeline, region:us, 102022 downloads, GGUF architecture gemma4, 131072 context length, gguf.total 11907350576, and gguf.totalFileSize 23832064640. The API totalFileSize matches gemma-4-12b-it-uncensored-f16.gguf, not the card-recommended Q4_K_M artifact selected for this profile." }, { "label": "zaakirio Gemma 4 12B IT Uncensored GGUF model card", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF/raw/32880562ac43cb589a85afb864309fdcaf486fae/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "selected_artifact", "sidecars", "serving" ], "notes": "The pinned card records Gemma licensing, base_model zaakirio/gemma-4-12b-it-uncensored, provenance from google/gemma-4-12B-it through Heretic ablation, llama.cpp serving, Q4_K_M as the recommended best size/quality balance, and usage examples with -hf zaakirio/gemma-4-12b-it-uncensored-GGUF:Q4_K_M. It also documents separate mmproj-gemma-4-12B-it-bf16.gguf and mtp-gemma-4-12b-it-uncensored.gguf sidecars." }, { "label": "zaakirio Gemma 4 12B IT Uncensored source config", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored/raw/5fdfef5f3aa9dd3c680a22078f2aa97b4a31c919/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "layer_types", "sliding_window", "max_context_tokens", "tie_word_embeddings" ], "notes": "The pinned full-precision source config records Gemma4UnifiedForConditionalGeneration, dtype bfloat16, tie_word_embeddings true, 48 text layers, eight full-attention layers, 40 sliding-attention layers, 1024-token sliding window, attention_k_eq_v true, one global KV head, 512 global head dimension, eight local KV heads, 256 local head dimension, 131072 max position embeddings, vision config, and audio config." }, { "label": "zaakirio Gemma 4 12B IT Uncensored GGUF linked-object HEAD checks", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF/tree/32880562ac43cb589a85afb864309fdcaf486fae", "source_type": "derived_calculation", "supports": [ "selected_artifact", "resident_weight_gb", "auxiliary_scope" ], "notes": "HEAD checks found gemma-4-12b-it-uncensored-Q4_K_M.gguf is 7381381760 bytes and is the card-recommended selected artifact. Sibling file sizes are Q2_K 4830147200 bytes, Q3_K_M 6087086720, Q4_K_S 7024046720, Q5_K_M 8547267200, Q6_K 9786020480, Q8_0 12669645440, f16 23832064640, mmproj-gemma-4-12B-it-bf16.gguf 175115584, and mtp-gemma-4-12b-it-uncensored.gguf 465109248." }, { "label": "zaakirio Gemma 4 12B IT Uncensored Q4_K_M GGUF range-read tensor index", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF/resolve/32880562ac43cb589a85afb864309fdcaf486fae/gemma-4-12b-it-uncensored-Q4_K_M.gguf", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "layers", "kv_heads", "head_dim", "sliding_window", "weight_format", "kv_adapter" ], "notes": "A 64MB range-read of the selected GGUF v3 header found 45 metadata entries and 667 tensors. The selected file is 7.381381760 GB, with header/tokenizer/alignment data starting tensor payloads at byte 15821952. Tensor spans total 7.365559808 GB across 11907350576 logical elements: token_embd.weight 0.825753600 GB, blk.* tensors 6.539789824 GB, output_norm.weight 0.000015360 GB, and rope_freqs.weight 0.000001024 GB. Tensor spans split into Q4_K 5.246484480 GB, Q6_K 2.115993600 GB, and F32 0.003081728 GB. Metadata/tokenizer/header/file overhead accounts for 0.015821952 GB. The header records general.architecture gemma4, block_count 48, context_length 131072, attention.head_count 16, layer KV head array with eight global layers using one KV head and 40 sliding layers using eight KV heads, key/value length 512, sliding_window 1024, shared_kv_layers 0, and no output.weight or mmproj/vision/audio/MTP tensor in the selected main file." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from live HF API metadata, pinned model card, pinned full-precision source config, HEAD checks for GGUF linked file sizes, and a direct GGUF header/tensor-index range read of the selected Q4_K_M artifact." }, "notes": "Use this profile for the card-recommended main Q4_K_M GGUF text artifact. Do not infer multimodal projector or MTP drafter residency or traffic unless those separate GGUF files are explicitly loaded by the workload." }, { "id": "zai-org--chatglm2-6b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/chatglm2-6b", "title": "ChatGLM2 6B FP16", "summary": "Audited memory-side text-decode bounds profile for the FP16 ChatGLM2 6B PyTorch checkpoint.", "model_family": "chatglm2-dense", "base_model_proof": { "base_model": "zai-org/chatglm2-6b", "relation": "base", "source": "Served config and PyTorch index in the target repo", "config_compatible": true, "notes": "This profile targets the served repo directly. The repo ships PyTorch shard files rather than safetensors, so exact resident bytes come from the PyTorch index metadata and exact swept/auxiliary splits are derived from the pinned config and checkpoint tensor names." }, "architecture": { "canonical_architecture_id": "chatglm2-6b", "max_context_tokens": 32768, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.243584032, "swept_params_b": 5.977245728, "auxiliary_resident_params_b": 0.266338304, "resident_weight_gb": 12.487168064, "swept_weight_gb": 11.954491456, "auxiliary_resident_weight_gb": 0.532676608, "resident_parameter_scope": "pytorch_index_total_size_fp16", "swept_parameter_scope": "transformer encoder layers, final layernorm, rotary inv_freq, and transformer.output_layer tensors derived from config tensor formulas", "auxiliary_scope": "transformer.embedding.word_embeddings.weight is resident for token lookup but is not swept as a full matrix for ordinary decode", "notes": "The PyTorch index records total_size 12487168064 bytes across 200 tensors. The checkpoint stores transformer.embedding.word_embeddings.weight separately from transformer.output_layer.weight because tie_word_embeddings is false. Ordinary decode sweeps the 28 transformer layers, final layernorm, tiny rotary inv_freq tensor, and output layer. The input embedding is resident-only for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records multi_query_attention true with multi_query_group_num 2 and kv_channels 128. The custom SelfAttention stores cached K/V before expanding multi-query heads to 32 query heads, so KV cache is charged for two K/V groups per layer." }, "notes": "ChatGLMModel uses a custom GLM transformer with multi-query attention, rotary embeddings, FP16 weights, 32768 sequence length, and cached decode via past_key_values." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "chatglm2-fp16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. PyTorch pickle/container overhead, activation traffic, attention kernels, and cache writes are outside this memory-side bound.", "notes": "The config records torch_dtype float16 and no KV-cache quantization scheme, so both KV storage and read traffic are charged at two bytes per scalar." }, "evidence": [ { "label": "ChatGLM2 6B model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/chatglm2-6b", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "revision", "downloads", "serving", "sequence_length", "multi_query_attention" ], "notes": "At repo SHA d2e2d91789248536a747d9ce60642a336444186c, the API records a public transformers repo with PyTorch checkpoint shards, custom_code, chatglm/glm/thudm tags, endpoints_compatible, and region:us. Current downloads are 413791. The model card describes ChatGLM2-6B as a bilingual chat model with 32K context and Multi-Query Attention." }, { "label": "ChatGLM2 6B config", "url": "https://huggingface.co/zai-org/chatglm2-6b/raw/d2e2d91789248536a747d9ce60642a336444186c/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records ChatGLMModel, torch_dtype float16, hidden_size 4096, 28 layers, 32 attention heads, kv_channels 128, multi_query_attention true, multi_query_group_num 2, ffn_hidden_size 13696, padded_vocab_size 65024, seq_length 32768, tie_word_embeddings false, use_cache true, RMSNorm, add_qkv_bias true, and no linear biases otherwise." }, { "label": "ChatGLM2 6B remote modeling code", "url": "https://huggingface.co/zai-org/chatglm2-6b/raw/d2e2d91789248536a747d9ce60642a336444186c/modeling_chatglm.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "swept_weight_gb" ], "notes": "Manual review found SelfAttention builds qkv_hidden_size as projection_size plus two multi-query K/V groups, stores kv_cache before expanding K/V to all query heads, and returns past_key_values through ChatGLMForConditionalGeneration. The model defines separate Embedding and transformer.output_layer modules, 28 GLMBlock layers, final_layernorm, and rotary_pos_emb.inv_freq." }, { "label": "ChatGLM2 6B PyTorch index and derived tensor formulas", "url": "https://huggingface.co/zai-org/chatglm2-6b/resolve/d2e2d91789248536a747d9ce60642a336444186c/pytorch_model.bin.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The PyTorch index records total_size 12487168064 bytes and 200 tensor names across seven shards. Config-derived FP16 tensor formulas give transformer.embedding.word_embeddings.weight = 65024 * 4096 * 2 = 0.532676608 GB resident-only. Each of the 28 layers has two 4096-element norms, qkv weight 4608 * 4096, qkv bias 4608, attention output weight 4096 * 4096, MLP h_to_4h weight 4096 * 27392, and MLP 4h_to_h weight 13696 * 4096, totaling 0.407921664 GB per layer. The swept set also includes final_layernorm 0.000008192 GB, transformer.output_layer.weight 0.532676608 GB, and the tiny rotary inv_freq tensor. Resident minus embedding gives exact swept traffic 11.954491456 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned config, remote modeling code, PyTorch shard index, and config-derived tensor formulas. Full PyTorch shards were not downloaded because the index plus deterministic config shapes provide the needed resident and swept byte evidence." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact PyTorch index resident size, explicit multi-query KV geometry, and an explicit input-embedding resident-only split." }, { "id": "zai-org--chatglm3-6b", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/chatglm3-6b", "title": "ChatGLM3 6B FP16", "summary": "Audited memory-side text-decode bounds profile for the FP16 ChatGLM3 6B safetensors checkpoint.", "model_family": "chatglm3-dense", "base_model_proof": { "base_model": "zai-org/chatglm3-6b", "relation": "base", "source": "Served config, remote modeling code, and safetensors index/header metadata in the target repo", "config_compatible": true, "notes": "This profile targets the served repo directly. The model card describes ChatGLM3-6B as the dialogue model in the ChatGLM3 series and says its base model is ChatGLM3-6B-Base, but the served repo itself contains the exact config, modeling code, and safetensors tensor headers needed for the bounds profile." }, "architecture": { "canonical_architecture_id": "chatglm3-6b", "max_context_tokens": 8192, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 6.243584032, "swept_params_b": 5.977245728, "auxiliary_resident_params_b": 0.266338304, "resident_weight_gb": 12.487168064, "swept_weight_gb": 11.954491456, "auxiliary_resident_weight_gb": 0.532676608, "resident_parameter_scope": "safetensors_header_stored_fp16", "swept_parameter_scope": "transformer encoder layers, final layernorm, rotary inv_freq, and transformer.output_layer tensors from direct safetensors headers", "auxiliary_scope": "transformer.embedding.word_embeddings.weight is resident for token lookup but is not swept as a full matrix for ordinary decode", "notes": "Range-read safetensors headers record 200 FP16 tensors totaling 6.243584032B stored parameters / 12.487168064 GB. The checkpoint stores transformer.embedding.word_embeddings.weight separately from transformer.output_layer.weight because tie_word_embeddings is false. Ordinary decode sweeps the 28 transformer layers, final layernorm, tiny rotary inv_freq tensor, and output layer. The input embedding is resident-only for ordinary decode." }, "kv_adapter": { "kind": "full_context", "layers": 28, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The config records multi_query_attention true with multi_query_group_num 2 and kv_channels 128. The custom SelfAttention stores cached K/V with two K/V groups before expanding multi-query heads to 32 query heads, so KV cache is charged for two K/V groups per layer." }, "notes": "ChatGLMModel uses a custom GLM transformer with multi-query attention, rotary embeddings, FP16 weights, 8192 sequence length, and cached decode via past_key_values." }, "serving": { "weight_format": "fp16", "weight_bytes_per_param": 2, "kv_store_format": "fp16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "fp16", "kv_read_bytes_per_scalar": 2, "runtime_format": "chatglm3-fp16-text-decode-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Safetensors container overhead, activation traffic, attention kernels, and cache writes are outside this memory-side bound.", "notes": "The config records torch_dtype float16 and no KV-cache quantization scheme, so both KV storage and read traffic are charged at two bytes per scalar." }, "evidence": [ { "label": "ChatGLM3 6B model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/chatglm3-6b", "source_type": "model_card", "supports": [ "repo", "license", "revision", "downloads", "serving", "sequence_length", "multi_query_attention" ], "notes": "At repo SHA e9e0406d062cdb887444fe5bd546833920abd4ac, the API records a public non-gated transformers repo with PyTorch and safetensors checkpoint shards, custom_code, chatglm/glm/thudm tags, endpoints_compatible, and region:us. Current downloads are 114420. The model card describes ChatGLM3-6B as the latest generation ChatGLM dialogue model, identifies ChatGLM3-6B-Base as its base model, and says the 32K variant is separate from this 8192-token repo." }, { "label": "ChatGLM3 6B config", "url": "https://huggingface.co/zai-org/chatglm3-6b/raw/e9e0406d062cdb887444fe5bd546833920abd4ac/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "kv_adapter", "serving", "tie_word_embeddings" ], "notes": "The config records ChatGLMModel, torch_dtype float16, hidden_size 4096, 28 layers, 32 attention heads, kv_channels 128, multi_query_attention true, multi_query_group_num 2, ffn_hidden_size 13696, padded_vocab_size 65024, seq_length 8192, tie_word_embeddings false, use_cache true, RMSNorm, add_qkv_bias true, and no linear biases otherwise." }, { "label": "ChatGLM3 6B remote modeling code", "url": "https://huggingface.co/zai-org/chatglm3-6b/raw/e9e0406d062cdb887444fe5bd546833920abd4ac/modeling_chatglm.py", "source_type": "manual_review", "supports": [ "architecture", "kv_adapter", "swept_weight_gb" ], "notes": "Manual review found SelfAttention builds qkv_hidden_size as projection_size plus two multi-query K/V groups, allocates K/V cache with num_multi_query_groups_per_partition when multi_query_attention is enabled, and returns past_key_values through ChatGLMForConditionalGeneration. The model defines separate Embedding and transformer.output_layer modules, 28 GLMBlock layers, final_layernorm, and rotary_pos_emb.inv_freq." }, { "label": "ChatGLM3 6B safetensors index and shard headers", "url": "https://huggingface.co/zai-org/chatglm3-6b/resolve/e9e0406d062cdb887444fe5bd546833920abd4ac/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "embedding_layout" ], "notes": "The safetensors index records total_size 12487168064 bytes and 200 tensor names across seven shards. Range-read shard headers found 200 FP16 tensors totaling 12.487168064 GB, matching the index total_size. transformer.embedding.word_embeddings.weight has shape [65024, 4096] and contributes 0.532676608 GB resident-only. transformer.output_layer.weight is stored separately with the same shape and remains in swept decode traffic. Transformer encoder layers plus final_layernorm, rotary_pos_emb.inv_freq, and output_layer total 11.954491456 GB. Linked-object HEAD checks resolved the seven safetensors shards to 12.487194904 GB total, leaving 26840 bytes of safetensors header/container overhead outside tensor payloads." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned config, remote modeling code, safetensors index, linked-object HEAD checks, direct range-read safetensors shard headers, and local scrape row." }, "notes": "This profile supersedes the scraped metadata estimate by using the exact safetensors header resident size, explicit multi-query KV geometry, and an explicit input-embedding resident-only split. The main difference from the existing ChatGLM2 6B profile is the served 8192-token sequence length instead of 32768." }, { "id": "zai-org--glm-4-1v-9b-thinking", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.1V-9B-Thinking", "title": "GLM 4.1V 9B Thinking BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 GLM-4.1V 9B Thinking multimodal repo.", "model_family": "glm4v-dense-multimodal", "base_model_proof": { "base_model": "zai-org/GLM-4-9B-0414", "relation": "finetune", "source": "Hugging Face model metadata, served config, base config comparison, Transformers GLM4V implementation review, and safetensors header review", "config_compatible": false, "notes": "The target repo records zai-org/GLM-4-9B-0414 as its base model and preserves the core text tensor geometry: 40 layers, hidden size 4096, intermediate size 13696, 32 attention heads, 2 KV heads, 128 head dimension, untied embeddings, and vocabulary size 151552. It changes the wrapper to glm4v_text, records 65536 max position embeddings instead of the base 32768, adds mRoPE metadata, and adds a resident GLM4V vision tower, so this profile audits the served repo directly." }, "architecture": { "canonical_architecture_id": "glm4v-9b-thinking", "max_context_tokens": 65536, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.292777472, "swept_params_b": 8.779522048, "auxiliary_resident_params_b": 1.513255424, "resident_weight_gb": 20.585554944, "swept_weight_gb": 17.559044096, "auxiliary_resident_weight_gb": 3.026510848, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode charges model.language_model.layers.*, model.language_model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.language_model.embed_tokens.weight and model.visual.* are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary text decode token", "notes": "Range-read safetensors headers record 704 BF16 tensors totaling 10292777472 stored parameters / 20.585554944 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.language_model.embed_tokens.weight and lm_head.weight tensors. The visual tower contributes 1.784996864 GB resident-only for ordinary text-token decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The GLM4V text attention implementation uses separate k_proj and v_proj tensors sized by num_key_value_heads times head_dim and updates standard key/value cache entries. No sliding-window or latent-cache adapter is declared, so Bounds Engine v1 charges full-context BF16 K and V streams for all text layers." }, "notes": "Dense GLM4V multimodal profile using the served text_config and direct safetensors header byte grouping. Vision execution and image/video prefill cost are outside ordinary text-decode bounds, but the loaded vision tower is included in resident memory." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-glm4v-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The served config records bfloat16 text dtype, and range-read safetensors headers record only BF16 tensors. KV cache bytes are charged as BF16." }, "evidence": [ { "label": "GLM-4.1V 9B Thinking API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.1V-9B-Thinking", "source_type": "model_card", "supports": [ "repo", "base_model_proof", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit 9e9a4c5e94f4a095c353f4152d520a2644a553b2, the public API records an MIT image-text-to-text repo with glm4v tags, base_model zai-org/GLM-4-9B-0414, endpoints_compatible, deploy:azure, region:us, and safetensors parameters BF16: 10292777472. Current downloads were 580375 when audited." }, { "label": "GLM-4.1V 9B Thinking config", "url": "https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking/raw/9e9a4c5e94f4a095c353f4152d520a2644a553b2/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "vision_residency", "serving", "base_model_proof" ], "notes": "The served config records Glm4vForConditionalGeneration, model_type glm4v, text_config model_type glm4v_text, BF16 text dtype, 40 text layers, hidden_size 4096, intermediate_size 13696, 32 attention heads, 2 KV heads, 65536 max position embeddings, tie_word_embeddings false, vocab_size 151552, and a 24-layer vision tower with hidden size 1536 and output hidden size 4096." }, { "label": "GLM-4 9B 0414 base config", "url": "https://huggingface.co/zai-org/GLM-4-9B-0414/raw/645b8482494e31b6b752272bf7f7f273ef0f3caf/config.json", "source_type": "manual_review", "supports": [ "base_model_proof" ], "notes": "Manual comparison found matching core text tensor geometry between the base config and target text_config, except the target uses glm4v_text, 65536 max position embeddings, mRoPE metadata, and a resident vision tower." }, { "label": "Transformers GLM4V implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.1/src/transformers/models/glm4v/modeling_glm4v.py", "source_type": "manual_review", "supports": [ "kv_adapter", "embedding_layout", "vision_residency" ], "notes": "Manual implementation review found Glm4vTextAttention k_proj and v_proj sized by num_key_value_heads times head_dim, cache updates for key/value states, Glm4vTextModel embed_tokens, Glm4vModel visual plus language_model modules, and a separate Glm4vForConditionalGeneration lm_head projection." }, { "label": "GLM-4.1V 9B Thinking safetensors headers", "url": "https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking/resolve/9e9a4c5e94f4a095c353f4152d520a2644a553b2/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "vision_residency", "embedding_layout" ], "notes": "Range reads of all four safetensors shard headers found 704 BF16 tensors totaling 10292777472 parameters / 20.585554944 GB, matching index total_size. model.visual.* tensors contribute 1.784996864 GB resident-only. model.language_model.embed_tokens.weight contributes 1.241513984 GB resident-only for ordinary decode. model.language_model.layers.*, model.language_model.norm.weight, and lm_head.weight total 17.559044096 GB of swept ordinary text-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, served config, base config comparison, Transformers GLM4V implementation review, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This profile is for ordinary text-token decode after any multimodal prefill. It includes the vision tower in residency but does not charge image/video encoder execution per generated text token." }, { "id": "zai-org--glm-4-5-air", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.5-Air", "title": "GLM 4.5 Air BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 GLM-4.5-Air repo.", "model_family": "glm4-moe", "architecture": { "canonical_architecture_id": "glm-4-5-air", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 220.93766144, "main_resident_weight_gb": 212.463000064, "auxiliary_resident_weight_gb": 8.474661376, "fixed_weight_gb": 13.149673984, "routed_expert_weight_gb": 1.55713536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary Transformers text decode through model.layers.0-45, model.norm, and lm_head, excluding input embedding lookup and auxiliary layer 46 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.46 tensors are resident for the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors in ordinary MoE layers are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes BF16 tensors with small F32 router correction biases. Expected-distinct routing is applied to the 128 uniform routed expert indexes across ordinary MoE layers 1-45." }, "kv_adapter": { "kind": "full_context", "layers": 46, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 46 ordinary decoder layers, 8 KV heads, 128 head dimension, and no sliding-window or compressed-cache setting. The audited Transformers glm4_moe attention path updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "The served config records 46 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The safetensors index exposes that auxiliary package as model.layers.46 tensors, which are kept resident but excluded from ordinary causal decode traffic." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000106600211, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and traffic GB fields account for the small F32 router correction-bias tensor set.", "notes": "The config records dtype bfloat16, and safetensors headers record 110,468,818,944 BF16 parameters plus 5,888 F32 parameters. KV cache bytes are charged from the audited expanded-cache Transformers path." }, "evidence": [ { "label": "GLM-4.5-Air model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5-Air", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit a24ceef6ce4f3536971efe9b778bdaa1bab18daa, the current API records a public MIT text-generation repo with glm4_moe tags, endpoints_compatible, region:us, and safetensors parameters BF16: 110468818944, F32: 5888, total: 110468824832. Current downloads were 398671 when audited. The model card describes GLM-4.5-Air as a 106B-total, 12B-active compact GLM-4.5 series MoE model and links the Transformers, vLLM, and SGLang implementations." }, { "label": "GLM-4.5-Air config", "url": "https://huggingface.co/zai-org/GLM-4.5-Air/raw/a24ceef6ce4f3536971efe9b778bdaa1bab18daa/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Glm4MoeForCausalLM, glm4_moe, bfloat16, 46 hidden layers, one next-token-prediction layer, first_k_dense_replace 1, hidden_size 4096, intermediate_size 10944, moe_intermediate_size 1408, 96 attention heads, 8 key/value heads, head_dim 128, 128 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab size 151552, rope_theta 1000000, and 131072 max position embeddings." }, { "label": "GLM-4.5-Air safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-4.5-Air/raw/a24ceef6ce4f3536971efe9b778bdaa1bab18daa/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 47 shards. Stored tensor payloads sum to 220.937661440 GB: 220.937637888 GB BF16 and 0.000023552 GB F32. model.embed_tokens.weight contributes 1.241513984 GB resident-only. The auxiliary model.layers.46 tensor group contributes 7.233147392 GB resident-only. Ordinary decode main resident tensors therefore sum to 212.463000064 GB. Routed expert tensors in ordinary layers 1-45 sum to 199.313326080 GB and divide exactly into 128 uniform expert groups of 1.557135360 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layer 0, attention, routers, shared experts, norms, and model.norm.weight, sums to 13.149673984 GB." }, { "label": "Transformers GLM4 MoE implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm4_moe/modeling_glm4_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found Glm4MoeModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 46 layers for this config. Glm4MoeAttention projects key_states and value_states, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx), supporting expanded full-context BF16 K/V cache charges for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-05", "notes": "Audited from current HF API metadata, model card, pinned served config, generation config, safetensors index, direct range-read safetensors shard headers, and the upstream Transformers glm4_moe runtime implementation." }, "notes": "This profile models ordinary text decode. It intentionally excludes the auxiliary next-token-prediction tensor package from non-speculative decode traffic and does not assume runtime-specific MTP/speculative serving." }, { "id": "zai-org--glm-4-5", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.5", "title": "GLM 4.5 BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 GLM-4.5 repo.", "model_family": "glm4-moe", "architecture": { "canonical_architecture_id": "glm-4-5", "max_context_tokens": 131072, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 716.675611392, "main_resident_weight_gb": 704.043794048, "auxiliary_resident_weight_gb": 12.631817344, "fixed_weight_gb": 32.116293248, "routed_expert_weight_gb": 4.19954688, "routed_experts": 160, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary Transformers text decode through model.layers.0-91, model.norm, and lm_head, excluding input embedding lookup and auxiliary layer 92 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.92 tensors are resident for the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors in ordinary MoE layers are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes BF16 tensors with small F32 router correction biases. Expected-distinct routing is applied to the 160 uniform routed expert indexes across ordinary MoE layers 3-91." }, "kv_adapter": { "kind": "full_context", "layers": 92, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 92 ordinary decoder layers, 8 KV heads, 128 head dimension, and no sliding-window or compressed-cache setting. The audited Transformers glm4_moe attention path updates standard key_states and value_states in past_key_values, so Bounds Engine v1 charges expanded BF16 K/V cache streams." }, "notes": "The served config records 92 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The safetensors index exposes that auxiliary package as model.layers.92 tensors, which are kept resident but excluded from ordinary causal decode traffic." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.0000000803710933, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and traffic GB fields account for the small F32 router correction-bias tensor set.", "notes": "The config records dtype bfloat16, and safetensors headers record 358,337,776,896 BF16 parameters plus 14,400 F32 parameters. KV cache bytes are charged from the audited expanded-cache Transformers path." }, "evidence": [ { "label": "GLM-4.5 model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit cbb2c7cfb52fa128a9660cb1a7a78e017899e115, the current API records a public MIT text-generation repo with glm4_moe tags, endpoints_compatible, region:us, and safetensors parameters BF16: 358337776896, F32: 14400, total: 358337791296. Current downloads were 144724 when audited. The model card describes GLM-4.5 as a 355B-total, 32B-active MoE model and documents 128K context deployment." }, { "label": "GLM-4.5 config", "url": "https://huggingface.co/zai-org/GLM-4.5/raw/cbb2c7cfb52fa128a9660cb1a7a78e017899e115/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Glm4MoeForCausalLM, glm4_moe, bfloat16, 92 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 5120, intermediate_size 12288, moe_intermediate_size 1536, 96 attention heads, 8 key/value heads, head_dim 128, 160 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab size 151552, rope_theta 1000000, and 131072 max position embeddings." }, { "label": "GLM-4.5 safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-4.5/raw/cbb2c7cfb52fa128a9660cb1a7a78e017899e115/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 93 shards. Stored tensor payloads sum to 716.675611392 GB: 716.675553792 GB BF16 and 0.000057600 GB F32. model.embed_tokens.weight contributes 1.551892480 GB resident-only. The auxiliary model.layers.92 tensor group contributes 11.079924864 GB resident-only. Ordinary decode main resident tensors therefore sum to 704.043794048 GB. Routed expert tensors in ordinary layers 3-91 sum to 671.927500800 GB and divide exactly into 160 uniform expert groups of 4.199546880 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, routers, shared experts, norms, and model.norm.weight, sums to 32.116293248 GB." }, { "label": "Transformers GLM4 MoE implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm4_moe/modeling_glm4_moe.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review from the existing GLM-4.5-Air audit found Glm4MoeModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 92 layers for this config. Glm4MoeAttention projects key_states and value_states, applies RoPE, and calls past_key_values.update(key_states, value_states, layer_idx), supporting expanded full-context BF16 K/V cache charges for ordinary decode." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, pinned served config, generation config, safetensors index, direct range-read safetensors shard headers, and the existing upstream Transformers glm4_moe runtime review." }, "notes": "This profile models ordinary text decode. It intentionally excludes the auxiliary next-token-prediction tensor package from non-speculative decode traffic and does not assume runtime-specific MTP/speculative serving." }, { "id": "zai-org--glm-4-5v", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.5V", "title": "Z.ai GLM-4.5V BF16 MoE", "summary": "Audited memory-side text-decode bounds profile for the BF16/F32 GLM-4.5V multimodal MoE checkpoint.", "model_family": "glm4v-moe-multimodal", "base_model_proof": { "base_model": "zai-org/GLM-4.5-Air-Base", "relation": "finetune", "source": "Hugging Face model card/API metadata, served config, and safetensors header grouping", "config_compatible": true, "notes": "The model card and API metadata record GLM-4.5-Air-Base as the base model. GLM-4.5V adds the GLM-V multimodal stack but keeps a standard autoregressive text-decode path through Glm4vMoeForConditionalGeneration with GLM-4.5-style MoE text layers." }, "architecture": { "canonical_architecture_id": "glm4v-moe-46l-128e", "max_context_tokens": 65536, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 215.42187776, "main_resident_weight_gb": 212.463000064, "auxiliary_resident_weight_gb": 2.958877696, "fixed_weight_gb": 13.149673984, "routed_expert_weight_gb": 1.55713536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary text decode through model.language_model excluding embed_tokens plus lm_head; routed expert traffic is the expected distinct 8-of-128 expert traffic from byte-uniform expert groups", "auxiliary_scope": "model.visual tensors and model.language_model.embed_tokens.weight are resident for multimodal packaging and token lookup but not swept as full matrices for each ordinary generated text token", "shared_expert_notes": "The config records n_shared_experts 1, and the checkpoint stores model.language_model.layers.*.mlp.shared_experts.* tensors outside the routed experts.* namespace. These shared expert tensors are always-on ordinary text traffic and are included in fixed_weight_gb.", "notes": "Range-read safetensors headers across all 46 shards matched the index total_size exactly: 215.421877760 GB across 18106 tensors. Routed expert tensors under model.language_model.layers.1-45.mlp.experts.* total 199.313326080 GB and divide exactly into 128 byte-uniform expert indexes of 1.557135360 GB. Fixed ordinary text/logit traffic is 13.149673984 GB, including attention, routers, shared experts, dense layer-0 MLP, norms, and lm_head.weight. Resident-only auxiliary bytes are 1.717363712 GB visual tensors plus 1.241513984 GB input embedding." }, "kv_adapter": { "kind": "full_context", "layers": 46, "kv_heads": 8, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The served config records 46 text layers, 8 KV heads, 128 head dimension, use_cache true, and no sliding-window or recurrent-state adapter. K and V projections are stored separately in the safetensors headers, so this profile charges separate K and V streams." }, "notes": "This profile models ordinary text decode after any image/video prefill. Vision encoder, projector, media preprocessing, video cache behavior, and multimodal prefill throughput are outside Bounds Engine v1." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "vllm-sglang-transformers-bf16-moe-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. The checkpoint stores nearly all tensors in BF16 plus tiny F32 router correction bias tensors. Activation traffic, kernels, expert scheduling, multimodal projector execution, and media prefill are outside Bounds Engine v1.", "notes": "The model card documents Transformers, vLLM, and SGLang serving. The memory-side profile uses the exact safetensors payload bytes rather than idealized parameter counts." }, "evidence": [ { "label": "Z.ai GLM-4.5V HF API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5V", "source_type": "derived_calculation", "supports": [ "repo", "base_model_proof", "downloads", "license", "pipeline", "total_params_b", "serving" ], "notes": "The current HF CLI/API response records commit ed47433b37111465ec527affaaddceff371bca04, MIT license, image-text-to-text pipeline, region:us, 134414 downloads, base_model zai-org/GLM-4.5-Air-Base, safetensors parameters BF16 107710927360 and F32 5760, and total 107710933120." }, { "label": "Z.ai GLM-4.5V config", "url": "https://huggingface.co/zai-org/GLM-4.5V/raw/ed47433b37111465ec527affaaddceff371bca04/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "serving", "tie_word_embeddings" ], "notes": "The config records Glm4vMoeForConditionalGeneration, text model_type glm4v_moe_text, BF16 dtype, 46 text layers, hidden size 4096, 96 attention heads, 8 KV heads, 128 head dimension, first_k_dense_replace 1, n_routed_experts 128, num_experts_per_tok 8, n_shared_experts 1, moe_intermediate_size 1408, max_position_embeddings 65536, use_cache true, tie_word_embeddings false, and a 24-layer visual tower." }, { "label": "Z.ai GLM-4.5V model card", "url": "https://huggingface.co/zai-org/GLM-4.5V/raw/ed47433b37111465ec527affaaddceff371bca04/README.md", "source_type": "model_card", "supports": [ "base_model_proof", "license", "pipeline", "runtime_format", "unsupported_exclusions" ], "notes": "The card states GLM-4.5V is based on GLM-4.5-Air, which it describes as a 106B-parameter text foundation model with 12B active parameters. It documents Transformers, vLLM, and SGLang serving, image/video/document/GUI usage, and recommends FA3 for SGLang multimodal inference. Those multimodal prefill and media-cache details are outside this ordinary text-decode profile." }, { "label": "Z.ai GLM-4.5V safetensors index and shard header audit", "url": "https://huggingface.co/zai-org/GLM-4.5V/raw/ed47433b37111465ec527affaaddceff371bca04/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb", "routed_experts", "weight_format" ], "notes": "The index records total_size 215421877760 bytes across 46 shards. Direct range reads of all shard headers found 18106 tensors totaling exactly 215.421877760 GB: 215.421854720 GB BF16 and 0.000023040 GB F32. The 46 linked shard files total 215.424400616 GB, leaving 0.002522856 GB of safetensors header/container bytes outside tensor payloads. model.language_model tensors total 212.463000064 GB, model.visual tensors total 1.717363712 GB, model.language_model.embed_tokens.weight is 1.241513984 GB, and lm_head.weight is a separate 1.241513984 GB tensor. Ordinary main text/logit resident bytes, defined as model.language_model excluding embed_tokens plus lm_head, total 212.463000064 GB. Routed expert tensors total 199.313326080 GB across layers 1-45 and divide exactly into 128 expert indexes of 1.557135360 GB. Fixed ordinary text/logit traffic totals 13.149673984 GB: attention 10.034094080 GB, shared experts 1.557135360 GB, dense layer-0 MLP 0.268959744 GB, routers 0.047208960 GB, norms 0.000761856 GB, and lm_head 1.241513984 GB." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF CLI/API metadata, pinned model card, served config, safetensors index, and direct header range reads for all 46 safetensors shards." }, "notes": "This profile is for ordinary text decode only. It deliberately excludes visual encoder/projector prefill throughput and treats input embeddings plus visual tensors as resident-only auxiliary bytes." }, { "id": "zai-org--glm-4-6v-flash", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.6V-Flash", "title": "GLM 4.6V Flash BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16 GLM-4.6V Flash multimodal repo.", "model_family": "glm4v-dense-multimodal", "architecture": { "canonical_architecture_id": "glm-4-6v-flash", "max_context_tokens": 131072, "weight_adapter": { "kind": "dense_resident_swept", "resident_params_b": 10.292777472, "swept_params_b": 8.779522048, "auxiliary_resident_params_b": 1.513255424, "resident_weight_gb": 20.585554944, "swept_weight_gb": 17.559044096, "auxiliary_resident_weight_gb": 3.026510848, "resident_parameter_scope": "safetensors_header_stored_bf16", "swept_parameter_scope": "ordinary text decode charges model.language_model.layers.*, model.language_model.norm.weight, and lm_head.weight", "auxiliary_scope": "model.language_model.embed_tokens.weight and model.visual.* are resident for token lookup and multimodal prefill but are not swept as full matrices for each ordinary text decode token", "notes": "Range-read safetensors headers record 704 BF16 tensors totaling 10292777472 stored parameters / 20.585554944 GB. The config marks tie_word_embeddings false and the checkpoint stores separate model.language_model.embed_tokens.weight and lm_head.weight tensors. The visual tower contributes 1.784996864 GB resident-only for ordinary text-token decode." }, "kv_adapter": { "kind": "full_context", "layers": 40, "kv_heads": 2, "head_dim": 128, "kv_scalar_multiplier": 2, "notes": "The GLM4V text attention implementation uses separate k_proj and v_proj tensors sized by num_key_value_heads times head_dim and updates standard key/value cache entries. No sliding-window or latent-cache adapter is declared, so Bounds Engine v1 charges full-context BF16 K and V streams for all text layers." }, "notes": "Dense GLM4V multimodal profile using the served text_config and direct safetensors header byte grouping. Vision execution and image/video prefill cost are outside ordinary text-decode bounds, but the loaded vision tower is included in resident memory." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2, "kv_store_format": "bf16", "kv_store_bytes_per_scalar": 2, "kv_read_format": "bf16", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-glm4v-memory-bound", "dequantization_notes": "No quantized weight representation is assumed for this BF16 repo.", "notes": "The served config records bfloat16 text dtype, and range-read safetensors headers record only BF16 tensors. KV cache bytes are charged as BF16." }, "evidence": [ { "label": "GLM-4.6V Flash API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.6V-Flash", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "resident_weight_gb", "weight_format" ], "notes": "At commit 411bb4d77144a3f03accbf4b780f5acb8b7cde4e, the public API records an MIT image-text-to-text repo with glm4v tags, endpoints_compatible, deploy:azure, region:us, and safetensors parameters BF16: 10292777472. Current downloads were 343891 when audited." }, { "label": "GLM-4.6V Flash model card", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash", "source_type": "model_card", "supports": [ "repo", "model_family", "pipeline", "max_context_tokens" ], "notes": "The model card identifies this repository as GLM-4.6V-Flash, a 9B lightweight GLM-V model for local deployment and low-latency applications. It describes 128K context training and documents vLLM, SGLang, and Transformers serving." }, { "label": "GLM-4.6V Flash config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json", "source_type": "config", "supports": [ "model_family", "layers", "kv_heads", "head_dim", "max_context_tokens", "vision_residency", "serving" ], "notes": "The served config records Glm4vForConditionalGeneration, model_type glm4v, text_config model_type glm4v_text, BF16 text dtype, 40 text layers, hidden_size 4096, intermediate_size 13696, 32 attention heads, 2 KV heads, 131072 max position embeddings, tie_word_embeddings false, vocab_size 151552, and a 24-layer vision tower with hidden size 1536 and output hidden size 4096." }, { "label": "Transformers GLM4V implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v5.0.0rc0/src/transformers/models/glm4v/modeling_glm4v.py", "source_type": "manual_review", "supports": [ "kv_adapter", "embedding_layout", "vision_residency" ], "notes": "Manual implementation review found Glm4vTextAttention k_proj and v_proj sized by num_key_value_heads times head_dim, cache updates for key/value states, Glm4vTextModel embed_tokens, Glm4vModel visual plus language_model modules, and a separate Glm4vForConditionalGeneration lm_head projection." }, { "label": "GLM-4.6V Flash safetensors headers", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/resolve/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "swept_weight_gb", "auxiliary_resident_weight_gb", "vision_residency", "embedding_layout" ], "notes": "Range reads of all four safetensors shard headers found 704 BF16 tensors totaling 10292777472 parameters / 20.585554944 GB. model.visual.* tensors contribute 1.784996864 GB resident-only. model.language_model.embed_tokens.weight contributes 1.241513984 GB resident-only for ordinary decode. model.language_model.layers.*, model.language_model.norm.weight, and lm_head.weight total 17.559044096 GB of swept ordinary text-decode traffic." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-06", "notes": "Audited from current HF API metadata, model card, served config, Transformers GLM4V implementation review, safetensors index, direct shard header range reads, and local scrape row." }, "notes": "This profile is for ordinary text-token decode after any multimodal prefill. It includes the vision tower in residency but does not charge image/video encoder execution per generated text token." }, { "id": "zai-org--glm-4-7-flash", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-4.7-Flash", "title": "GLM 4.7 Flash BF16/F32", "summary": "Audited memory-side text-decode bounds profile for the mixed BF16/F32 GLM-4.7-Flash repo.", "model_family": "glm4-moe-lite", "architecture": { "canonical_architecture_id": "glm-4-7-flash", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 62.442983168, "main_resident_weight_gb": 59.252405248, "auxiliary_resident_weight_gb": 3.19057792, "fixed_weight_gb": 3.686265856, "routed_expert_weight_gb": 0.868220928, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary Transformers text decode through layers 0-46, excluding input embedding lookup and auxiliary layer 47 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.47 tensors are resident for the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived bytes are used because the package mixes BF16 tensors with small F32 router correction biases. Expected-distinct routing is applied to the 64 uniform routed expert indexes across layers 1-46." }, "kv_adapter": { "kind": "full_context", "layers": 47, "kv_heads": 20, "head_dim": 256, "kv_scalar_multiplier": 2, "notes": "Although the config exposes MLA projection fields, the audited upstream Transformers implementation expands cached key_states and value_states before past_key_values.update. Bounds Engine v1 therefore charges full BF16 K and V cache streams for all 47 ordinary decoder layers." }, "notes": "The served config records 47 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 47 tensors are kept resident but excluded from ordinary causal decode traffic." }, "serving": { "weight_format": "mixed_bf16_f32", "weight_bytes_per_param": 2.000000192687802, "kv_store_format": "expanded_bf16_key_value_cache", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-f32-expanded-kv-memory-bound", "dequantization_notes": "No quantized weight representation is assumed. Exact resident and traffic GB fields account for the small F32 router correction-bias tensor set.", "notes": "The config records dtype bfloat16, and safetensors headers record 31,221,485,568 BF16 parameters plus 3,008 F32 parameters. KV cache bytes are charged from the audited expanded-cache Transformers path, not from the latent projection rank alone." }, "evidence": [ { "label": "GLM-4.7-Flash model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.7-Flash", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 7dd20894a642a0aa287e9827cb1a1f7f91386b67, the API reports an MIT text-generation repo with glm4_moe_lite tags and safetensors parameters BF16: 31221485568, F32: 3008, total: 31221488576. The model card describes GLM-4.7-Flash as a 30B-A3B MoE model and documents local serving through vLLM, SGLang, and Transformers main branches." }, { "label": "GLM-4.7-Flash config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records Glm4MoeLiteForCausalLM, glm4_moe_lite, bfloat16, 47 hidden layers, one next-token-prediction layer, first_k_dense_replace 1, hidden_size 2048, 20 attention heads, 20 key/value heads, qk_nope_head_dim 192, qk_rope_head_dim 64, v_head_dim 256, kv_lora_rank 512, 64 routed experts, 4 experts per token, 1 shared expert, tie_word_embeddings false, and 202752 max position embeddings." }, { "label": "GLM-4.7-Flash safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 48 shards. Stored tensors sum to 62.442983168 GB: 62.442971136 GB BF16 and 0.000012032 GB F32. The input embedding tensor contributes 0.63438848 GB resident-only. The auxiliary layer 47 tensor group contributes 2.55618944 GB resident-only. Ordinary decode main resident tensors therefore sum to 59.252405248 GB. Routed expert tensors in layers 1-46 sum to 55.566139392 GB and divide exactly into 64 uniform expert groups of 0.868220928 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layer 0, attention, gates, norms, routers, and shared experts, sums to 3.686265856 GB." }, { "label": "Transformers GLM4 MoE Lite implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm4_moe_lite/modeling_glm4_moe_lite.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope" ], "notes": "Manual review found Glm4MoeLiteModel instantiates range(config.num_hidden_layers), so the ordinary decoder stack has 47 layers. Glm4MoeLiteAttention computes compressed_kv but expands it to key_states and value_states before past_key_values.update, so this profile charges expanded full-context BF16 K/V cache rather than a latent compressed cache." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, range-read safetensors shard headers, and the upstream Transformers glm4_moe_lite runtime implementation." }, "notes": "This profile models ordinary text decode. It intentionally does not assume runtime-specific MLA cache compression for vLLM or SGLang without direct implementation evidence." }, { "id": "zai-org--glm-5-1-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-5.1-FP8", "title": "GLM 5.1 FP8", "summary": "Audited memory-side text-decode bounds profile for the mixed FP8/BF16/F32 GLM-5.1-FP8 repo.", "model_family": "glm-moe-dsa", "architecture": { "canonical_architecture_id": "glm-5-1", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 756.162687872, "main_resident_weight_gb": 744.226889472, "auxiliary_resident_weight_gb": 11.9357984, "fixed_weight_gb": 19.274211072, "routed_expert_weight_gb": 2.8318464, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary Transformers text decode through layers 0-77, excluding input embedding lookup and auxiliary layer 78 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.78 tensors are resident in the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the repo stores most matrices as F8_E4M3, selected embeddings and projections as BF16, and scale/correction tensors as F32. Expected-distinct routing is applied to the 256 uniform routed expert indexes across sparse layers 3-77." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 192 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 256 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 1, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "DSA indexer key cache component from index_head_dim 128 stored per ordinary decoder layer." } ], "notes": "The upstream Transformers GLM MoE DSA implementation computes compressed MLA projections internally but expands key_states and value_states before past_key_values.update, and DynamicIndexedLayer stores a separate indexer key cache. Bounds Engine v1 therefore charges expanded BF16 K/V cache streams plus indexer key state, not a latent MLA cache or runtime-specific sparse-read optimization." }, "notes": "The served config records 78 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 78 tensors are kept resident but excluded from ordinary causal decode traffic." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0029879743844663, "kv_store_format": "expanded_bf16_key_value_cache_plus_dsa_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache_plus_dsa_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-expanded-kv-dsa-indexer-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads, BF16 embeddings/projections/norms, and F32 scale/correction tensors from safetensors headers. Dequantization, activation traffic, speculative decoding, and compute overhead are outside Bounds Engine v1.", "notes": "The config records dtype bfloat16 with fp8 e4m3 block quantized weights. KV and indexer cache bytes are charged as BF16 from the audited Transformers path." }, "evidence": [ { "label": "GLM-5.1-FP8 model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.1-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit f396cf805182f4ca10fa675e1a99815b3ca384db, the API reports a public MIT text-generation repo with transformers, safetensors, glm_moe_dsa, endpoints_compatible, and fp8 tags. Safetensors parameter metadata reports F32: 45904480, BF16: 2114950400, F8_E4M3: 751749169152, total: 753910024032." }, { "label": "GLM-5.1-FP8 config", "url": "https://huggingface.co/zai-org/GLM-5.1-FP8/raw/f396cf805182f4ca10fa675e1a99815b3ca384db/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records GlmMoeDsaForCausalLM, glm_moe_dsa, dtype bfloat16, 78 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 6144, intermediate_size 12288, moe_intermediate_size 2048, 64 attention heads, 64 key/value heads, q_lora_rank 2048, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, qk_head_dim 256, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab_size 154880, max_position_embeddings 202752, and fp8 e4m3 block quantization. Manual comparison found these ordinary text-decode geometry fields match the earlier audited GLM-5-FP8 config." }, { "label": "GLM-5.1-FP8 safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-5.1-FP8/resolve/f396cf805182f4ca10fa675e1a99815b3ca384db/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 142 shards. Stored tensors sum exactly to the index total_size of 756.162687872 GB: 751.749169152 GB F8_E4M3, 4.2299008 GB BF16, and 0.18361792 GB F32. The input embedding tensor contributes 1.90316544 GB resident-only, and auxiliary layer 78 contributes 10.03263296 GB resident-only. Ordinary decode main resident tensors therefore sum to 744.226889472 GB. Routed expert tensors in layers 3-77 sum to 724.9526784 GB and divide exactly into 256 uniform expert groups of 2.8318464 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, gates, norms, routers, and shared experts, sums to 19.274211072 GB." }, { "label": "Transformers GLM MoE DSA implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GlmMoeDsaModel instantiates range(config.num_hidden_layers) and loops through self.layers[: self.config.num_hidden_layers], so ordinary decode uses 78 layers. The pre-trained class ignores unexpected model.layers.78 tensors on load. GlmMoeDsaAttention computes compressed_kv, expands it into key_states with qk_nope_head_dim plus qk_rope_head_dim and value_states with v_head_dim, then calls past_key_values.update. GlmMoeDsaIndexer separately calls past_key_values.update_indexer with index_head_dim keys." }, { "label": "Transformers cache implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/cache_utils.py", "source_type": "manual_review", "supports": [ "kv_adapter" ], "notes": "Manual review found DynamicIndexedLayer extends DynamicLayer with indexer_keys, and update_indexer concatenates tensors of shape [batch_size, seq_len, index_head_dim]. StaticIndexedLayer similarly preallocates [batch_size, max_cache_len, index_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from the served config, API metadata, range-read safetensors shard headers, config comparison with GLM-5-FP8, and the pinned upstream Transformers GLM MoE DSA runtime and cache implementation." }, "notes": "This profile models ordinary text decode. It intentionally does not assume runtime-specific MLA cache compression, FP8 KV cache, DSA sparse-read savings, or speculative decoding speedups for vLLM or SGLang without direct implementation evidence." }, { "id": "zai-org--glm-5-2-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-5.2-FP8", "title": "GLM 5.2 FP8", "summary": "Audited memory-side text-decode bounds profile for the mixed FP8/BF16/F32 GLM-5.2-FP8 repo with IndexShare DSA.", "model_family": "glm-moe-dsa", "architecture": { "canonical_architecture_id": "glm-5-2", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 755.617140416, "main_resident_weight_gb": 743.681342016, "auxiliary_resident_weight_gb": 11.9357984, "fixed_weight_gb": 18.728663616, "routed_expert_weight_gb": 2.8318464, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary Transformers text decode through layers 0-77, excluding input embedding lookup and auxiliary layer 78 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.78 tensors are resident in the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the repo stores most matrices as F8_E4M3, selected embeddings and projections as BF16, and scale/correction tensors as F32. Expected-distinct routing is applied to the 256 uniform routed expert indexes across sparse layers 3-77." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 192 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 256 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 21, "kv_heads": 1, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "DSA IndexShare indexer key cache component. The served config has 21 full indexer layers and 57 shared layers, so shared layers reuse top-k indices and do not run update_indexer." } ], "notes": "The upstream Transformers GLM MoE DSA implementation still expands key_states and value_states before past_key_values.update on all 78 ordinary layers. GLM-5.2 adds IndexShare through config.indexer_types, so Bounds Engine v1 charges expanded BF16 K/V streams for 78 layers plus BF16 DSA indexer key state for the 21 full indexer layers only." }, "notes": "The served config records 78 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 78 tensors are kept resident but excluded from ordinary causal decode traffic. GLM-5.2 raises max_position_embeddings to 1048576 and records an IndexShare indexer pattern of 21 full and 57 shared ordinary layers." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.002975071473079, "kv_store_format": "expanded_bf16_key_value_cache_plus_indexshare_dsa_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache_plus_indexshare_dsa_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-expanded-kv-indexshare-dsa-indexer-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads, BF16 embeddings/projections/norms, and F32 scale/correction tensors from safetensors headers. Dequantization, activation traffic, speculative decoding, and compute overhead are outside Bounds Engine v1.", "notes": "The config records dtype bfloat16 with fp8 e4m3 block quantized weights. KV and full-indexer cache bytes are charged as BF16 from the audited Transformers path." }, "evidence": [ { "label": "GLM-5.2-FP8 model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.2-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit ba978f7d347eaf65d22f1a86833408afdb953541, the API reports a public MIT text-generation repo with transformers, safetensors, glm_moe_dsa, endpoints_compatible, fp8, and region:us tags. Safetensors parameter metadata reports F32: 45872560, BF16: 2103729152, F8_E4M3: 751226191872, total: 753375793584. The model card describes GLM-5.2 as a 1M-token-context successor to GLM-5.1 with IndexShare." }, { "label": "GLM-5.2-FP8 config", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/raw/ba978f7d347eaf65d22f1a86833408afdb953541/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records GlmMoeDsaForCausalLM, glm_moe_dsa, dtype bfloat16, 78 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 6144, intermediate_size 12288, moe_intermediate_size 2048, 64 attention heads, 64 key/value heads, q_lora_rank 2048, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, qk_head_dim 256, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab_size 154880, max_position_embeddings 1048576, fp8 e4m3 block quantization, index_topk_freq 4, index_skip_topk_offset 3, and indexer_types with 21 full layers and 57 shared layers. Manual comparison found the ordinary text-decode geometry fields match GLM-5-FP8 and GLM-5.1-FP8 except for max_position_embeddings and IndexShare indexer scheduling." }, { "label": "GLM-5.2-FP8 safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/resolve/ba978f7d347eaf65d22f1a86833408afdb953541/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 141 shards. Stored tensors sum exactly to the index total_size of 755.617140416 GB: 751.226191872 GB F8_E4M3, 4.207458304 GB BF16, and 0.18349024 GB F32 across 118629 tensors. The input embedding tensor contributes 1.90316544 GB resident-only, and auxiliary layer 78 contributes 10.03263296 GB resident-only. Ordinary decode main resident tensors therefore sum to 743.681342016 GB. Routed expert tensors in layers 3-77 sum to 724.9526784 GB and divide exactly into 256 uniform expert groups of 2.8318464 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, gates, norms, routers, shared experts, and the GLM-5.2 full-indexer tensors, sums to 18.728663616 GB." }, { "label": "Transformers GLM MoE DSA configuration", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/configuration_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "indexshare_schedule" ], "notes": "Manual review found GlmMoeDsaConfig derives qk_head_dim as qk_nope_head_dim plus qk_rope_head_dim, derives mlp_layer_types as three dense layers plus sparse MoE layers, derives layer_types as deepseek_sparse_attention for every ordinary hidden layer, and derives indexer_types from index_topk_freq/index_skip_topk_offset when the saved config does not provide an explicit list. GLM-5.2 provides an explicit 78-entry indexer_types list with 21 full and 57 shared layers." }, { "label": "Transformers GLM MoE DSA implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GlmMoeDsaModel instantiates range(config.num_hidden_layers) and loops through self.layers[: self.config.num_hidden_layers], so ordinary decode uses 78 layers. GlmMoeDsaAttention computes compressed_kv, expands it into key_states with qk_nope_head_dim plus qk_rope_head_dim and value_states with v_head_dim, then calls past_key_values.update on every layer. It skips GlmMoeDsaIndexer when config.indexer_types[layer_idx] is shared, otherwise calls past_key_values.update_indexer for full indexer layers." }, { "label": "Transformers cache implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/cache_utils.py", "source_type": "manual_review", "supports": [ "kv_adapter" ], "notes": "Manual review found DynamicIndexedLayer extends DynamicLayer with indexer_keys, and update_indexer concatenates tensors of shape [batch_size, seq_len, index_head_dim]. StaticIndexedLayer similarly preallocates [batch_size, max_cache_len, index_head_dim]. GLM-5.2 only calls update_indexer from full indexer layers, so this profile charges indexer state for 21 layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-03", "notes": "Audited from the served config, API metadata, model card, range-read safetensors shard headers, config comparison with GLM-5-FP8 and GLM-5.1-FP8, and the pinned upstream Transformers GLM MoE DSA runtime/config/cache implementation." }, "notes": "This profile models ordinary text decode. It intentionally does not assume runtime-specific MLA cache compression, FP8 KV cache, DSA sparse-read savings beyond the config-proven IndexShare indexer-cache reduction, or speculative decoding speedups for vLLM or SGLang without direct implementation evidence." }, { "id": "zai-org--glm-5-2", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-5.2", "title": "GLM 5.2 BF16", "summary": "Audited memory-side text-decode bounds profile for the BF16/F32 GLM-5.2 repo with IndexShare DSA.", "model_family": "glm-moe-dsa", "architecture": { "canonical_architecture_id": "glm-5-2", "max_context_tokens": 1048576, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 1506.659919872, "main_resident_weight_gb": 1484.850912768, "auxiliary_resident_weight_gb": 21.809007104, "fixed_weight_gb": 15.972097536, "routed_expert_weight_gb": 5.737807872, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_bf16_f32", "traffic_scope": "ordinary Transformers text decode through layers 0-77, excluding input embedding lookup and auxiliary layer 78 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.78 tensors are resident in the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the repo stores almost all tensors as BF16 and a tiny set of tensors as F32. Expected-distinct routing is applied to the 256 uniform routed expert indexes across sparse layers 3-77." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 192 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 256 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 21, "kv_heads": 1, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "DSA IndexShare indexer key cache component. The served config has 21 full indexer layers and 57 shared layers, so shared layers reuse top-k indices and do not run update_indexer." } ], "notes": "The upstream Transformers GLM MoE DSA implementation still expands key_states and value_states before past_key_values.update on all 78 ordinary layers. GLM-5.2 records an IndexShare indexer pattern, so Bounds Engine v1 charges expanded BF16 K/V streams for 78 layers plus BF16 DSA indexer key state for the 21 full indexer layers only." }, "notes": "The served config records 78 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 78 tensors are kept resident but excluded from ordinary causal decode traffic. GLM-5.2 raises max_position_embeddings to 1048576 and records an IndexShare indexer pattern of 21 full and 57 shared ordinary layers." }, "serving": { "weight_format": "bf16", "weight_bytes_per_param": 2.00000005165333, "kv_store_format": "expanded_bf16_key_value_cache_plus_indexshare_dsa_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache_plus_indexshare_dsa_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-bf16-expanded-kv-indexshare-dsa-indexer-memory-bound", "dequantization_notes": "The memory-side bound charges stored BF16 matrix payloads and tiny F32 tensors from safetensors headers. Activation traffic, speculative decoding, and compute overhead are outside Bounds Engine v1.", "notes": "The config records dtype bfloat16. KV and full-indexer cache bytes are charged as BF16 from the audited Transformers path." }, "evidence": [ { "label": "GLM-5.2 model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.2", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit b4734de4facf877f85769a911abafc5283eab3d9, the API reports a public MIT text-generation repo with transformers, safetensors, glm_moe_dsa, endpoints_compatible, eval-results, region:us, and conversational tags. Safetensors parameter metadata reports BF16: 753329921024, F32: 19456, total: 753329940480. Live downloads during audit were 231218. The model card describes GLM-5.2 as a 1M-token-context successor to GLM-5.1 with IndexShare." }, { "label": "GLM-5.2 config", "url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records GlmMoeDsaForCausalLM, glm_moe_dsa, dtype bfloat16, 78 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 6144, intermediate_size 12288, moe_intermediate_size 2048, 64 attention heads, 64 key/value heads, q_lora_rank 2048, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, qk_head_dim 256, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab_size 154880, max_position_embeddings 1048576, index_topk_freq 4, index_skip_topk_offset 3, and indexer_types with 21 full layers and 57 shared layers. Manual comparison found the ordinary text-decode geometry fields match the audited GLM-5.2-FP8 sibling except for stored weight precision." }, { "label": "GLM-5.2 safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-5.2/resolve/b4734de4facf877f85769a911abafc5283eab3d9/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 282 shards. Stored tensors sum exactly to the index total_size of 1506.659919872 GB: 1506.659842048 GB BF16 and 0.000077824 GB F32 across 59585 tensors. The input embedding tensor contributes 1.903165440 GB resident-only, and auxiliary layer 78 contributes 19.905841664 GB resident-only. Ordinary decode main resident tensors therefore sum to 1484.850912768 GB. Routed expert tensors in layers 3-77 sum to 1468.878815232 GB and divide exactly into 256 uniform expert groups of 5.737807872 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, gates, norms, routers, shared experts, and the GLM-5.2 full-indexer tensors, sums to 15.972097536 GB. The shared-expert tensors contribute 5.737807872 GB and are included in fixed traffic. Ordinary full-indexer tensors contribute 0.393619968 GB." }, { "label": "Transformers GLM MoE DSA configuration", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/configuration_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "indexshare_schedule" ], "notes": "Manual review found GlmMoeDsaConfig derives qk_head_dim as qk_nope_head_dim plus qk_rope_head_dim, derives mlp_layer_types as three dense layers plus sparse MoE layers, derives layer_types as deepseek_sparse_attention for every ordinary hidden layer, and derives indexer_types from index_topk_freq/index_skip_topk_offset when the saved config does not provide an explicit list. GLM-5.2 provides an explicit 78-entry indexer_types list with 21 full and 57 shared layers." }, { "label": "Transformers GLM MoE DSA implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GlmMoeDsaModel instantiates range(config.num_hidden_layers) and loops through self.layers[: self.config.num_hidden_layers], so ordinary decode uses 78 layers. GlmMoeDsaAttention computes compressed_kv, expands it into key_states with qk_nope_head_dim plus qk_rope_head_dim and value_states with v_head_dim, then calls past_key_values.update on every layer. It skips GlmMoeDsaIndexer when config.indexer_types[layer_idx] is shared, otherwise calls past_key_values.update_indexer for full indexer layers." }, { "label": "Transformers cache implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/cache_utils.py", "source_type": "manual_review", "supports": [ "kv_adapter" ], "notes": "Manual review found DynamicIndexedLayer extends DynamicLayer with indexer_keys, and update_indexer concatenates tensors of shape [batch_size, seq_len, index_head_dim]. StaticIndexedLayer similarly preallocates [batch_size, max_cache_len, index_head_dim]. GLM-5.2 only calls update_indexer from full indexer layers, so this profile charges indexer state for 21 layers." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-07", "notes": "Audited from current HF API metadata, model card, pinned served config, direct range-read safetensors shard headers, comparison with the audited GLM-5.2-FP8 sibling, and the pinned upstream Transformers GLM MoE DSA runtime/config/cache implementation." }, "notes": "This profile models ordinary text decode. It intentionally does not assume runtime-specific MLA cache compression, FP8 KV cache, DSA sparse-read savings beyond the config-proven IndexShare indexer-cache reduction, or speculative decoding speedups for vLLM or SGLang without direct implementation evidence." }, { "id": "zai-org--glm-5-fp8", "version": "1.0.0", "schema_version": "1.0.0", "status": "audited", "repo": "zai-org/GLM-5-FP8", "title": "GLM 5 FP8", "summary": "Audited memory-side text-decode bounds profile for the mixed FP8/BF16/F32 GLM-5-FP8 repo.", "model_family": "glm-moe-dsa", "architecture": { "canonical_architecture_id": "glm-5", "max_context_tokens": 202752, "weight_adapter": { "kind": "moe_distinct_experts_exact", "resident_weight_gb": 756.162687872, "main_resident_weight_gb": 744.226889472, "auxiliary_resident_weight_gb": 11.9357984, "fixed_weight_gb": 19.274211072, "routed_expert_weight_gb": 2.8318464, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "routing_model": "uniform_expected_distinct", "resident_parameter_scope": "safetensors_header_stored_fp8_bf16_f32", "traffic_scope": "ordinary Transformers text decode through layers 0-77, excluding input embedding lookup and auxiliary layer 78 next-token-prediction tensors", "auxiliary_scope": "model.embed_tokens.weight and model.layers.78 tensors are resident in the checkpoint but not swept as full matrices for each ordinary generated token", "shared_expert_notes": "The config records n_shared_experts 1. Shared expert tensors are included in fixed_weight_gb because they are always-on MoE-layer traffic.", "notes": "Header-derived stored bytes are used because the repo stores most matrices as F8_E4M3, selected embeddings and projections as BF16, and scale/correction tensors as F32. Expected-distinct routing is applied to the 256 uniform routed expert indexes across sparse layers 3-77." }, "kv_adapter": { "kind": "layered_kv", "components": [ { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded key cache component from qk_nope_head_dim 192 plus qk_rope_head_dim 64 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 64, "head_dim": 256, "kv_scalar_multiplier": 1, "notes": "Expanded value cache component from v_head_dim 256 across all ordinary decoder layers." }, { "kind": "full_context", "layers": 78, "kv_heads": 1, "head_dim": 128, "kv_scalar_multiplier": 1, "notes": "DSA indexer key cache component from index_head_dim 128 stored per ordinary decoder layer." } ], "notes": "The upstream Transformers GLM MoE DSA implementation computes compressed MLA projections internally but expands key_states and value_states before past_key_values.update, and DynamicIndexedLayer stores a separate indexer key cache. Bounds Engine v1 therefore charges expanded BF16 K/V cache streams plus indexer key state, not a latent MLA cache or runtime-specific sparse-read optimization." }, "notes": "The served config records 78 ordinary hidden layers plus one num_nextn_predict_layers auxiliary layer. The layer 78 tensors are kept resident but excluded from ordinary causal decode traffic." }, "serving": { "weight_format": "fp8", "weight_bytes_per_param": 1.0029879743844663, "kv_store_format": "expanded_bf16_key_value_cache_plus_dsa_indexer", "kv_store_bytes_per_scalar": 2, "kv_read_format": "expanded_bf16_key_value_cache_plus_dsa_indexer", "kv_read_bytes_per_scalar": 2, "runtime_format": "transformers-fp8-expanded-kv-dsa-indexer-memory-bound", "dequantization_notes": "The memory-side bound charges stored F8_E4M3 matrix payloads, BF16 embeddings/projections/norms, and F32 scale/correction tensors from safetensors headers. Dequantization, activation traffic, speculative decoding, and compute overhead are outside Bounds Engine v1.", "notes": "The config records dtype bfloat16 with fp8 e4m3 block quantized weights. KV and indexer cache bytes are charged as BF16 from the audited Transformers path." }, "evidence": [ { "label": "GLM-5-FP8 model card and API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5-FP8", "source_type": "model_card", "supports": [ "repo", "license", "pipeline", "total_params_b", "serving" ], "notes": "At commit 4f96cc5eec29dcee5d6ded54f7ffe889438f9516, the API reports an MIT text-generation repo with transformers, safetensors, glm_moe_dsa, and fp8 tags. Safetensors parameter metadata reports F32: 45904480, BF16: 2114950400, F8_E4M3: 751749169152, total: 753910024032. The model card describes GLM-5 as scaling GLM-4.5 from 355B/32B active to 744B/40B active and integrating DeepSeek Sparse Attention, with local deployment guidance for vLLM, SGLang, KTransformers, Transformers, and xLLM." }, { "label": "GLM-5-FP8 config", "url": "https://huggingface.co/zai-org/GLM-5-FP8/raw/4f96cc5eec29dcee5d6ded54f7ffe889438f9516/config.json", "source_type": "config", "supports": [ "model_family", "routed_experts", "routed_experts_per_token", "shared_experts_per_token", "max_context_tokens", "kv_adapter", "serving" ], "notes": "The config records GlmMoeDsaForCausalLM, glm_moe_dsa, dtype bfloat16, 78 hidden layers, one next-token-prediction layer, first_k_dense_replace 3, hidden_size 6144, intermediate_size 12288, moe_intermediate_size 2048, 64 attention heads, 64 key/value heads, q_lora_rank 2048, kv_lora_rank 512, qk_nope_head_dim 192, qk_rope_head_dim 64, qk_head_dim 256, v_head_dim 256, index_head_dim 128, index_n_heads 32, index_topk 2048, 256 routed experts, 8 experts per token, 1 shared expert, tie_word_embeddings false, vocab_size 154880, max_position_embeddings 202752, and fp8 e4m3 block quantization." }, { "label": "GLM-5-FP8 safetensors index and shard headers", "url": "https://huggingface.co/zai-org/GLM-5-FP8/raw/4f96cc5eec29dcee5d6ded54f7ffe889438f9516/model.safetensors.index.json", "source_type": "derived_calculation", "supports": [ "resident_weight_gb", "main_resident_weight_gb", "auxiliary_resident_weight_gb", "fixed_weight_gb", "routed_expert_weight_gb" ], "notes": "Safetensors headers were range-read across all 142 shards. Stored tensors sum exactly to the index total_size of 756.162687872 GB: 751.749169152 GB F8_E4M3, 4.2299008 GB BF16, and 0.18361792 GB F32. The input embedding tensor contributes 1.90316544 GB resident-only, and auxiliary layer 78 contributes 10.03263296 GB resident-only. Ordinary decode main resident tensors therefore sum to 744.226889472 GB. Routed expert tensors in layers 3-77 sum to 724.9526784 GB and divide exactly into 256 uniform expert groups of 2.8318464 GB. Non-expert ordinary decode traffic, including lm_head.weight, dense layers 0-2, attention, gates, norms, routers, and shared experts, sums to 19.274211072 GB." }, { "label": "Transformers GLM MoE DSA implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py", "source_type": "manual_review", "supports": [ "kv_adapter", "ordinary_decode_scope", "auxiliary_resident_scope" ], "notes": "Manual review found GlmMoeDsaModel instantiates range(config.num_hidden_layers) and loops through self.layers[: self.config.num_hidden_layers], so ordinary decode uses 78 layers. The pre-trained class ignores unexpected model.layers.78 tensors on load. GlmMoeDsaAttention computes compressed_kv, expands it into key_states with qk_nope_head_dim plus qk_rope_head_dim and value_states with v_head_dim, then calls past_key_values.update. GlmMoeDsaIndexer separately calls past_key_values.update_indexer with index_head_dim keys." }, { "label": "Transformers cache implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/cache_utils.py", "source_type": "manual_review", "supports": [ "kv_adapter" ], "notes": "Manual review found DynamicIndexedLayer extends DynamicLayer with indexer_keys, and update_indexer concatenates tensors of shape [batch_size, seq_len, index_head_dim]. StaticIndexedLayer similarly preallocates [batch_size, max_cache_len, index_head_dim]." } ], "review": { "reviewed_by": "local-frontier-profile-review", "reviewed_at": "2026-07-02", "notes": "Audited from the served config, model card/API metadata, range-read safetensors shard headers, and the pinned upstream Transformers GLM MoE DSA runtime and cache implementation." }, "notes": "This profile models ordinary text decode. It intentionally does not assume runtime-specific MLA cache compression, FP8 KV cache, DSA sparse-read savings, or speculative decoding speedups for vLLM or SGLang without direct implementation evidence." } ] };